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

Destructors are one of the many pieces of the puzzle that is the C++ memory management.

[11.1] What's the deal with destructors?

FAQ: A destructor, abbreviated "dtor", is a "you're about to die" member function. It is used to release resources, for example, semaphores. Most frequently the resource is memory allocated by new; the destructor frees it with delete.

FQA: Yep, C++ has no garbage collection, which makes "memory" a manually managed "resource". Die-hard C++ programmers firmly believe that RAII - acquiring resources in constructors and releasing them in destructors - is the silver bullet for resource management. The problem with this approach is that many objects can point to the same piece of data, but only one of them is the "owner", and when the owner dies, its destructor promptly blows the piece of data to even smaller little pieces.

The most common way around this is copying. For example, you load a large file representing a 3D object model or an XML document or something. You find an interesting leaf on the 3D tree model or in the XML tree or whatever. Now you want to release the memory of the whole model so that you can free some space for processing the part. You need to either convince the model object that it doesn't own the stuff in your leaf, or you have to copy that stuff. The latter is way easier (try to convince an owner such as std::list to let a couple of nodes go out of its jurisdiction).

Which is one reason that can cause garbage collection to outperform manual memory management. But of course with garbage collection you'll never master your debugging tools to an extent anywhere near the expertise you'll reach with manual memory management (you'll have 10 times less bugs to practice on). It's up to you to choose the right tradeoff.

There are two common counter arguments: not all resources are memory, and not all systems can afford garbage collection. Well, surely more than 95% of resources are memory, and surely more than 95% of C++ code runs in systems which can afford garbage collection. These arguments are yet another evidence that C++ is about theory, not practice. And the resources which are not memory and must be released deterministically usually aren't handled very well be C++ destructors anyway. That's because a C++ destructor can't handle errors (it has no return value and it can't throw exceptions).

[11.2] What's the order that local objects are destructed?

FAQ: In reverse order of construction - the stuff declared last is destroyed first.

FQA: Which makes sense, but are you sure you want to write code that depends on this?

[11.3] What's the order that objects in an array are destructed?

FAQ: In reverse order of construction - the last element is destroyed first.

FQA: Which sort of makes sense, but you surely don't want to write code that depends on that one, do you?

By the way, don't forget to use delete[], not just delete, to free arrays allocated with new[], otherwise the stupid thing will not bother to check the number of elements pointed by your pointer and some of your destructors won't get called (actually, in theory it could be worse - the behavior is undefined).

[11.4] Can I overload the destructor for my class?

FAQ: No. Destructors never have parameters or return values. And you're not supposed to call destructors explicitly, so you couldn't use parameters or return values anyway.

FQA: Why do you want to overload destructors? Do you like C++ overloading? Are you sure? But the lack of return values is a pity - no way to handle errors. Let's hope they won't happen, shall we?

[11.5] Should I explicitly call a destructor on a local variable?

FAQ: Don't do that! The destructor will be called again at the end of scope. And the second call will do nasty things, like crashing your program or something.

FQA: There are standard questions to ask in these cases, like "if this syntax is never useful, why does it compile"? Anyway, you can call a destructor on a local variable, but you have to make sure you call a constructor after that and before the object goes out of scope. For example:

AMiserableFileClass file("f1.txt");
//use file... now we want to close it, but there's no close() method, so:
//open another file
new (&file) AMiserableFileClass("f2.txt");

AMiserableFileClass is a miserable file class because it has no close method, so you might feel the need to close it in the ugly way above. Try to avoid this, because most people won't understand it, and they shouldn't, because there are better things to do than fiddling with the many ugly bits of C++.

[11.6] What if I want a local to "die" before the close } of the scope in which it was created? Can I call a destructor on a local if I really want to?

FAQ: No, no, no! See the next FAQ for a simple solution. But don't call the destructor!

FQA: If you ever get into the situation of promoting an especially disgusting product, such as the C++ programming language, there are better ways to handle it than get all excited. You can try and find a less disgusting product to center your money-making around, or you can just relax.

"Don't call the destructor". There has to be some reason for the destructor call to compile, doesn't it? Perhaps sharing it with us could help calm down.

[11.7] OK, OK already; I won't explicitly call the destructor of a local; but how do I handle the above situation?

FAQ: Now you're talking! Here's a tip - you can simply surround the object with an artificial block:

  AMiserableFileClass file("f1.txt");
} //the file object dies here

FQA: Is this ugly code supposed to be better than the ugly code calling a destructor and than a constructor? On a level, this version is more cryptic (at least it's easy to see what gets called when in that other ugly piece of code). But the truth is that for many actual C++ programmers the "artificial blocks" (isn't all code "artificial"?) are more readable.

[11.8] What if I can't wrap the local in an artificial block?

FAQ: If you absolutely have to, do something like adding a close method to AMiserableFileClass which closes the file in its destructor. So you can achieve the effect of the destructor call without calling the destructor. Which is a taboo, get it?

The FAQ also points out how hard it is to write a close function so that the destructor doesn't try to close a closed file.

FQA: If you can change AMiserableFileClass, it's better than using ugly code to work around its deficiencies. But where's the FAQ's brave new reuse-oriented world now? What about all those classes with stable interfaces you can't change? Seriously, it can happen with classes not available in source form.

I sincerely believe that the not-so-appetizing "call the destructor, then call a constructor" method is legal C++. If this is indeed so, I fail to understand the problem with mentioning it in the answer.

As to the problem of having destructors and close functions respect each other - one way around this is to avoid non-trivial constructors and destructors and always use init and close functions. Next, you can replace C++ classes with their metaphysical "encapsulation" with forward-declared C structs, which can actually yield stable binary interfaces and save recompilation. Next, you can replace C++ with C (where execution time matters) or with one of the many sane, safe languages (where development time matters). There, doesn't it feel much better now?

[11.9] But can I explicitly call a destructor if I've allocated my object with new?

FAQ: You can't, unless the object was allocated with placement new. Objects created by new must be deleted, which does two things (remember them): calls the destructor, then frees the memory.

FQA: Translation: delete is a way to explictly call a destructor, but it also deallocates the memory. You can also call a destructor without deallocating the memory. It's ugly and useless in most cases, but you can do that.

Questions like "What exactly does delete do?" probably cross the fine line separating between the questions everyone claiming to know C++ should be able to answer and the questions identifying people with too much C++-related garbage in their brains. People can be quite productive by simply calling new and delete without thinking too much about the two steps involved, so not knowing about the two steps is probably no big deal. The most interesting thing about the two steps is that the idea of having the code using the class managing the memory of its objects is the main reason C++ code is recompiled so frequently.

[11.10] What is "placement new" and why would I use it?

FAQ: There are many uses for placement new. For example, you can allocate an object in a particular location, you can pass the pointer to this location to placement new like this: C* p = new(place) C(args);

Don't use this unless you really care where the object lives (say, you have a memory-mapped hardware device like a timer, and you want to have a Timer object at the particular location).

Beware! It is your job to make sure that the address where you allocate the object is properly aligned and that you have enough space for the object. You are also responsible for calling the destructor with p->~C();.

FQA: Size is relatively easy to deal with because we have sizeof, alignment can be more painful. Here's an advice: if you care where your objects are allocated, don't make them objects of classes with constructors and destructors and stuff. This way, you can create memory pools with code like C pool[N];, which takes care of alignment and "type safety". But if the class C has a constructor, it's going to get called N times by this statement (slow and stupid), and if you want someone to use placement new with the pool, you'll have to call the destructors after the constructors finish running (slow, stupid and cryptic). Or you can use char pool[N*sizeof(C);] with platform-specific additions to handle alignment, and then you won't be able to easily inspect the pool in a debugger (the object pool has the wrong type), etc.

And you have to be out of your mind to call placement new when you deal with memory-mapped hardware. Do you realize that the constructor is going to directly modify the state of the hardware? Even if you get this right (yes, there are ways to get this wrong, too boring to enumerate here), this is one very unmaintainable way to write this kind of code. If you think that's "intuitive", think again. What is the constructor doing - "creating" the timer? Come on, it's hardware, it was already physically there. Oh, it "initializes" it? So why don't use a function with "init" in its name instead of a "constructor"?

You have enough trouble thinking about the semantics of the hardware interface, so why would anyone want to add the complexity of C++ to the problem?

At least the FAQ has finally disclosed the top secret information about explicitly calling destructors.

[11.11] When I write a destructor, do I need to explicitly call the destructors for my member objects?

FAQ: No, they are called implicitly in the reverse order of their declaration in the class.

FQA: Pay attention to the details. If your member is a pointer to something allocated with new, the pointer's destructor, which is a purely metaphysical entity doing nothing physically observable, is called, but it's your job to call delete on the pointer, which calls the destructor of the pointed object. The pointed object is technically not a member of your class (the pointer is). The difference between the intuitive feeling that "it's part of the class" and the formal definition of the term "member" is one reason making const less than useful (is changing the object pointed by a member "changes" the object itself or not?).

Experienced C++ programmers find this natural since the C++ approach to these issues is relatively uniform. If you have the tough luck of using C++ a lot, this point is obvious (but you wouldn't ask the question in the first place).

There's a nifty thing called "garbage collection" that's been around for nearly half a century. The run time system automatically figures out which objects are not pointed by anything and collects this garbage. Check it out, it's quite nice.

[11.12] When I write a derived class's destructor, do I need to explicitly call the destructor for my base class?

FAQ: No, this is done implicitly, in the reverse order of the appearance of the classes in the inheritance list (but virtual inheritance complicates matters), and after the destruction of the class members.

FQA: The FAQ says that if someone is relying on more complicated details of the finalization order, you'll need information outside the scope of the FAQ. The remark probably refers to useful information like locations of mental institutions and drug prescriptions.

Seriously, a Usenet language FAQ is already a place normally visited only by language lawyers. If you need more obscure information to get your job done, the only legitimate reason is that you're writing a C++ compiler. If that's the case, and you feel miserable, cheer up - it could be worse. For example, imagine the misery of the people who'll use your compiler!

[11.13] Should my destructor throw an exception when it detects a problem?

FAQ: THAT IS DANGEROUS! Which is discussed in a FAQ about exceptions.

FQA: It shouldn't. And you can't return an error code either. However, on many systems you can send an e-mail with the word "BUMMER" to your support e-mail address.

If you want your destructor to detect problems, make it a close function.

[11.14] Is there a way to force new to allocate memory from a specific memory area?

FAQ: Oh, yessssss! Pools! Pools, pools, pools, pools, pools...

Please forgive me this time. I can't summarize what the FAQ is saying. It's almost as long as all the previous answers. And it's totally out of control. If you feel like it, fasten your seat belt, grab a barf bag, follow the link and knock yourself out.

FQA: Oh, nooooooo! Don't do that!

I know that your system has more than one kind of memory and/or you have real time requirements and you can't use a traditional malloc implementation. But trust me, you're going to hate the day you've heard about those pools. Pools have arbitrary size limits (at most 20 objects of this, at most 30 objects of that...). Everybody is going to have a paranoia attack and set the limits to huge values, and then you're out of memory, and then you start thinking about the "right" limits, and it turns out that it's extremely hard to figure that out, because seemingly independent modules have related memory requirements (say, they handle mutually exclusive situations so you'd like them to use the same memory, but how?), and then you need some hairy logic to compute those values on a per-configuration basis, which means building several versions of the program, and maybe creating scripts for computing the values at build time, and you're going to hate the day you've heard about those pools.

If you're absolutely sure you can't create a single allocator managing your memory, at least don't allocate objects of classes with constructors from pools, use C-like structs instead. If you refuse to give up and admit that you're doing a pretty low-level thing and instead want to keep pretending that you're "programming in the problem domain", and for some reason you think C++ classes help you with that - go ahead. The punishments for your crime are discussed throughout the FQA.

Copyright © 2007-2009 Yossi Kreinin
revised 17 October 2009