Slashdot Mirror
Experiences w/ Garbage Collection and C/C++?
dberger queries: "Java has helped garbage collection enter the mainstream programmer's life - but it's certainly not new or unique to java. There have been (and are) several ways to add garbage collection to C/C++ - the most active seeming to be Hans Boehm's free libgc. I'm curious if any of the Slashdot crowd has used this (or any other) C++ garbage collector in non-trivial commercial applications. If so - what were your experiences? If not, why not? (Before you ask, yes - I know that GC isn't the only difference between C++ and Java, but 'automagic memory management' is certainly part of Java's marketing luster)"
Bjarne Stroustrup, the creator of C++, has this to say on garbage collection:
Clearly, if your code has new operations, delete operations, and pointer arithmetic all over the place, you are going to mess up somewhere and get leaks, stray pointers, etc. This is true independently of how conscientious you are with your allocations: eventually the complexity of the code will overcome the time and effort you can afford. It follows that successful techniques rely on hiding allocation and deallocation inside more manageable types.
He goes on to give detailed examples and recommendations on how to avoid using garbage collection.
void* operator new(size_t n) {
return GC_malloc(n);
}
void operator delete(void* p) {
}
You can also mix collected memory with uncollected memory, but we really don't see the point. This way we can still have descructors which do useful things but the actual memory clean up is left to the garbage collector. Of course, as we write more and more new code we leave our deletes and our destructors out, and eventually we'll go through and remove them all. Until then, we can disable the garbage collector just by #if 0ing these lines out.
How we know is more important than what we know.
The Boehm-Weiser (BW) collector is not as portable as we had hoped. There are a number of platforms we wanted to run on where it just doesn't run at all. Relatively small changes to the target runtime can create a need to port it all over again. OpenBSD, in particular, was an ongoing hassle until we abandoned BW. Hans, I hasten to add, was quite encouraging, but he simply doesn't have time to adequately support the collector.
The BW collector doesn't work in our application. OpenCM has a few very large objects. For reasons we don't really understand, this tends to cause a great deal of garbage retention when running the BW collector. Enough so that the OpenCM server crashed a lot when using it. Please note that this was NOT a bug involving falsely retained pointers, as later experience showed.
Conservative collectors are actually too conservative. If you are willing to make very modest changes in your source code as you design the app, there prove to be very natural places in the code for collection, and the resulting collector is quite efficient.
Independent of the collector, we also hacked together an exceptions package. This was also the right thing to do, but it's easy to trip over it in certain ways. The point of mentioning this is that once you do exceptions the pointer tracking becomes damned near hopeless and you essentially have to go to GC.
I think the way to say this is: exceptions + GC reduces your error handling code by a lot. Instead of three lines of error check on every procedure call, the error checking is confined to logical recovery points in the program, and you don't have to mess around simulating multiple return values in order to return a result code in parallel with the actually intended return value.
To provide malloc pluggability, we implemented an explicit free operation. This lets us interoperate compatibly with other libraries and do leak detection. Turns out to be very handy in lots of ways.
Hybrid storage management works very well. For example, our diff routine explicitly frees some of its local storage (example) [Sorry -- this link will go stale within the next few weeks because the OpenCM web interface will change in a way that makes it obsolete. If the link doesn't work for you, try looking for the same file in .../DEV/opencm/...] This is actually quite wonderful, as it lets us build certain libraries to be GC compatible without being GC dependent. One of the challenges in using a GC'd runtime in a library is compatibility with an enclosing application that doesn't use GC. We haven't tried it yet, but it looks like our gcmalloc code will handle this.
Eventually, we gave up on the BW collector and wrote our own. Our collector is conceptually very similar to the collector that Keith Packard built for Nickle, though we've since built from there. A variant of the Nickle collector is also used as a debugging leak tracer for X11.
The OpenCM GC system is reasonably well standalone. We need to document it, but others might want to look at it when we cut our next release.
On the whole, I'ld say that GC for this app was definitely the right thing to do. Once you get into object caches it becomes very hard to locate all of the objects and decide when to free them. We were able to use a conservative approach with no real hassle, and heap size is fairly well bounded by the assisted GC approach we took.
On the other hand, I would not recommend a pure conservative collector for a pro
Jonathan S. Shapiro (The EROS Guy)
Garbage collection has costs:
- The obvious: CPU & memory overhead for the checking and tracking. I can't comment on the amount here, but it is a generalized solution, so you forego the optimization opportunities that you'd otherwise have.
- The subtle: Memory allocation can become a major bottleneck in multithreaded systems. Garbage collection has similar issues.
- The irritating: you don't know when your destructors are called.
Another way: Smart Pointers. They're simple wrappers around the types that act like pointers, but they can make sure your objects live as long as you need and no longer. The big trick is knowing which kind of smart pointer you want.
- Reference Counting Smart Pointer (RCSP for short): this type of smart pointer will keep of how many RCSPs are pointing to the same object. It'll delete the object when the last RCSP is destroyed. A good one is the boost shared_ptr. Available for free from www.boost.org. This type is great for general use.
- Owning Smart Pointer (OSP): this type is specialized for those cases when the refcnt is never more than 1. When you assign one OSP (a) to another (b), the new OSP (a) gets ownership of the referred object, and the old one (b) is automatically set to null. When an OSP that isn't set to null is destroyed, it deletes the object it owns. It's great for parameter passing, return values, and objects you want dead at the end of the current scope, even if there's an exception. The STL comes with auto_ptr, which works this way.
You can use an RCSP wherever you can use an OSP, but not the other way around. The STL containers are a great example.
Sure it's not as easy as 'allocate and forget,' but you won't have the (sometimes very costly) expense of full-blown garbage collection.
Also, you can optimize your smart pointers for individual types (through template specialization). A great example is to give the no-longer-needed object back to a pool for later reuse.
This is really a quick, quick overview. For the meat & potatoes, go read Effective STL by Scott Meyers.
I've tried really hard to be fair & polite. There's probably still a bias, but I'm really trying!!
Care about electronic freedom? Consider donating to the EFF!
So please tell me what
typedef vector::const_iterator Iter;
(or rather vector::const_iterator) is supposed to mean. I suppose vector is a templated class, but how does ::const_iterator come up with a type name -- I thought :: either references a static field or a class member function?
And what is the deal with the sort(,) as a free-standing function? Following OO principles, shouldn't the vector object v know how to sort itself with a call to v.sort()? And what the heck is this const_iterator type anyway that you can do ++ and * on it -- looks an awful lot like a pointer -- oops, I forgot, you can overload ++ and * to make "safe" operations on what are really objects look like "dangerous" pointer operations which the C/C++ community is in custom of using.
In principle, all the stuff done in Java and perhaps in scripting languages could all be done elegantly and expressively in C++ if us mere mortals ever figure out how to use the darned thing. But there is a kind of uniformity to Java (all object variables being GC'd heap references, collection and iterator types working with generic Object's that we cast to what the object is and rely on runtime type checking, don't worry-be happy allocation of these objects where we grab towels from the rack and leave them on the bathroom floor for the hotel maid to pick up) that simply feels more comfortable.
C++ is the music of Bach: elegant, mathematical, intricate, and expressive, but most musicians performing in front of audiences don't understand it and it is played as a dull jumble and mishmash and audiences gaze at Bach stuck at the beginning of a recital as a chore to get through. Java is the music of Mozart: simplified, standardized, predictable, and economical, but musicians of this era understand it and play it with gusto, and audiences love it because it sound so happy and makes them feel uplifted.
A program variable is either a global variable, a stack variable, a class variable or an instance variable. Global and stack variables are held in lists. Class and instance variables are kept inside objects.
Every class object has a global variable that always refers to it.
Any object that is not, and that can not become referenced (directly or indirectly) by a global or stack program variable is garbage.
Each object has a 'not-garbage' flag.
For each global and stack variable, if the referenced object is not marked not-garbage, mark the referenced object as not-garbage, and recurse for that objects contained variables.
Delete all objects that are not marked not-garbage.
There are a few more twists, like handling return values on the stack, but this algorithm correctly handles self-referencing objects no matter the complexity.
Does everything include nothing?
There is another indirect cost pointed out by Linus Torvalds in a lengthy post to the gcc mailining list. The executive summary is that (he thinks that) memory that is not to be used anymore should be freed immediately. Otherwise, the data in there will keep lying around in the data cache. Also, he claims that explicit ref-counting gives you advantages for optimization: Assume you have to make some modifications to a data structure, but you don't want other parts of the program to see the modifications. Without ref-counting, you have to copy all the data structure before modifying it. With ref-couting, you can omit the copying if you are the only one with access to the data structure.
And finally, he thinks that GC makes it too easy to write pointer-chasing-heavy code---as that kind of code is bad for cache behaviour all the time.
It is an ongoing discussion whether GC really has that bad effects on performance of GCC. But Linus Torvalds seems to have very good points. (And some of them certainly cannot be taken into account in a "GC cost is less than hand-written memory management"-paper.)
I've used a garbage collection system in a C project before and it works surprisingly well. The problem with GC in C though is that it is possible and legal to,
o allocate memory
o write the pointer to a disk
o lose the pointer in memory
o read the pointer back off the disk,
o make use of the pointer
With all GC strategies I'm aware of, by the time you read the pointer from the disk the memory may well have been freed.
I'm not saying that this style of programming is a generally good idea but it is used in certain, specialised situations and therefore not suitable for a garbage collecting language.
The salient points:
Destructors are not Called
If an object is allocated in collectible memory, then its destructor will not be called when the object is collected. Therefore, destructors are pretty much useless and your code must be designed to work without them.
Actually, if your object derives from class gc_cleanup, then its destructor will be called. However, due to the handling of cleanup functions in the BDW collector, cycles of such objects will never be collected. For this reason, I don't use gc_cleanup much.
Allocating Collectible Memory
By default, C++ allocates objects in the "malloc" heap. The BDW collector maintains a separate heap. In effect, there are four types of memory:
"Scannable" refers to the property that objects in the heap are scanned for pointers. "Collectible" refers to the property that objects in the heap will be deallocated if no further references are found.
These four memory types are an issue when you interact with STL and third-party class libraries. By default, STL uses the malloc heap. If you want, say, a std::vector in collectible memory, then you need to write an allocator to get it. The most recent versions of the collector come with such a beast; the version I started using did not.
Similarly, std::string is reference-counted, and in the malloc heap. Here, rather than using an allocator to force it into the collectible heap, I wrote my own lightweight GCString class, which stores the string as an immutable object, and relies on the collector for cleanup.
Third-party class libraries such as ANTLR may use reference-counted objects; you need to bridge between GC and non-GC applications carefully.
I pointed out that Stroustrup's example shows the expressive power of C++, but there is a big "huh?" factor of reading the code on account that many of us mere mortals are not rehearsed in the use of templates and STL, and there was something to be said for Java and GC, not just for safety but for simplicity of expression and code reading by maintenance programmers.
Yours is one of three comments to my remarks, answering questions I had raised, disagreeing with some points, agreeing with others, but otherwise engaging in a reasoned discussion of the merits of C++ and its advanced features (templates and STL). But I get moderated down and flagged "troll" -- oh well.
With the first example I belive the point is that the string and vector classes will clean up themselves when they go out of scope (when their destructors are called). STL is very helpful especially when supplemented with the Boost libraries.
A few years ago, I used the Boehme GC when writing a pair of compilers (Verilog/VHDL) in C++. I was very happy with the result, since it was rare for GC even to get called at all. It was also surprising how much simpler code gets when you don't have to worry about deleting objects.
There's a really good book about everything you ever needed to know about garbage collection. Although most of the book deals with garbage collection techniques in general, it has two complete chapters devoted to implementing and using garbage collectors in C and C++ and which ones you should use depending on your application needs.
iow, it's really worthwhile to think about memory management issues, who owns memory, etc, and not just for the free() call, but for the design itself. It pays back a lot to think of these things, and use GC for particular cases that can benefit.
I agree with this, actually. I've seen many cases where having to think about object lifetimes has given me clearer insights into the problem domain and into the design, and resulted in better, tighter, cleaner and more maintainable code than would have been the case otherwise.
However, I've also seen some systems where the cleanest, most flexible and most maintainable design made keeping track of object lifetimes a real bitch. So while I wholeheartedly agree that being forced to think about memory management can improve the overall quality of most designs, there are circumstances where having to manage memory manually forces you to choose an inferior application architecture. In those cases, GC is a *huge* win, since choosing the right structure is the single most important thing you can do to facilitate both implementation and maintenance.
In addition, there are those cases where the code just doesn't matter that much; and GC is certainly an aid to getting the job done quickly, because it removes a large class of concerns from the programmer.
I'm not necessarily a huge fan of GC, sometimes it's great, sometimes it sucks -- it's just another tool. But I had to respond in this thread because there are a lot of misconceptions about GC (and the alternatives, like refcounting smart pointers).
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.