I think the only time you should need a separate Release method (instead of just a destructor) is if:
- you are implementing your own reference counting (e.g. like a COM object does, paired with AddRef), or
- your object holds onto system resources and you want your object to be able release those important system resources before it goes out of scope or is destroyed (such as a File object releasing a file handle, a texture object releasing the underlying GPU resource, something like that).
It sounds like there is some conflation of independent concepts.
A resource might be acquired and released; the interface for this is pretty much arbitrary and irrelevant.
A handle to a resource should obey RAII principles (assuming C++) and automatically acquire the resource during construction and release the resource during destruction.
An object might have handles to many resources, in which case it should still use RAII design. A handle should never be to more than one resource. A resource itself might be wrapped in a number of ways, but it should always be interacted with via a handle.
This solves all the problems that have been mentioned so far :-)
Personally I use CComPtr for (d3d) objects and only set them to null in the destructor, no releasing/ release calls needed.
A stripped example:
#include <atlbase.h>
class CD3d
{
public:
CComPtr<IDirect3DDevice9> mD3ddev; // smart pointer, first LPDIRECT3DDEVICE9
// etc
CD3d::CD3d()
{
mD3ddev = NULL; // smart pointer
// etc
CD3d::~CD3d()
{
// Smart pointers
mD3ddev = NULL;
// etc
This way I never have memory leaks.
Not 100% sure though if it's really necessary to set them to NULL / nullptr. If not, you could save yourself making your own constructor/destructor (rule of 3?) and let the compiler do it automatically for you.
The problem as I see it is by having a Release() method, you are then implying that instances of your class can exist in two states - formed and released. After you call Release(), your object is still around and can still be operated upon, so your internals of the class and your users then need to code around this, checking the state of the object during any operations.
One of the strongest principles of encapsulation is class invariants - if I have a std::string object, I am guaranteed from the outside to have an object containing a valid string or the empty string. I don't have to check isValid() every time.
In an ideal world (in which I accept we are not) it is better to use object lifetime to model the lifetime of the resource being modelled, then your objects can and need only exist in a formed state.
Thank you guys, I have a clear vision now, and think I'll get rid of Release methods for a while, since I use raw pointers and no ref counts. Long life to destructors.
The problem as I see it is by having a Release() method, you are then implying that instances of your class can exist in two states - formed and released. After you call Release(), your object is still around and can still be operated upon, so your internals of the class and your users then need to code around this, checking the state of the object during any operations.
One of the strongest principles of encapsulation is class invariants - if I have a std::string object, I am guaranteed from the outside to have an object containing a valid string or the empty string. I don't have to check isValid() every time.
No, that isn't what I am saying at all.
You want something that can be created almost instantly. Think of it as an empty string, or an empty vector, or an empty hash table. That doesn't mean it is ill-formed, that doesn't mean it is invalid, that doesn't mean RAII doesn't apply. It is well-formed and complete. It just means it is currently empty.
Consider the ever-popular epic work, the GameObject class. It serves as the base class of everything scriptable and a game world can have tens of thousands or even hundreds of thousands of them. A continuous world might have many million game objects in it.
Usually this thing grows to be MASSIVE. It isn't just a single transformation with position and orientation. It is usually hooked up to lots of components. it can be associated with models and textures and animations and shaders. It can be visible to the player, or not. It can be active and alive and important to the player, or it can be on the other side of the game universe just sitting around waiting for the player to walk across the map. Each child class can add their own details and containers and resource needs. It gets connected and disconnected to systems everywhere in the game.
When you start up the game you generally don't want every game object to load every detail immediately when it is constructed. For a large game the process could take five, ten or twenty minutes. For a continuous world game, you could easily exhaust all your memory.
When I say "Give me an array of 50 game objects", the absolute last thing I want to have happen is to cascade into 50 meshes loading, 50-300 textures loading, 500+ animations loading, 30+ shaders loading, tons of audio loading, followed by objects self-registering inside the world and the spatial trees, appending themselves to the render lists, including itself in the pathfinder, and so on. Such a process might mean thousands of calls to the disk as file names are turned into data streams and each stream individually pulled from a slow-spinning platter, followed by hundreds of calls into some of the most expensive systems of the game. No! I want 50 completely empty game objects that I can use for my own purposes immediately.
The key detail is that you don't do all of those things during construction and destruction. Instead, you do it at a time that you can control, a time that is appropriate. You can replace the very expensive models and textures and animations with simple proxies when they are off screen; they can be reloaded in moments when the player approaches them. You can remove proximity checks when the player is on the other side of the world, the land mine won't be stepped on when the player is on the other side of the world.
I am not saying to not initialize your objects. Far from it. Everything should be in a valid state. What I am saying is that "valid state" does not mean "requires 47MB of resources", but can also mean "empty", "ready for delayed loading", "proxied", and potentially much more. Careful resource management is very important in games.
frob - Yes, I understand that you weren't suggesting uninitialized objects. What I am saying is that you can model what you are talking about using composition and object lifetime, then you don't need to worry about the "empty" state if the empty state only needs to exist in order to accommodate delayed loading for example.
If the empty state is not a conceptual feature of the object, but an implementation detail to allow for delayed loading, this can be expressed via having the object contained in a proxy then the acquisition (delayed) and freeing of the actual object can be modelled using the object lifetime, just as a member of a proxy object.
The proxy can then handle releasing its owned resource, possibly in a generic way and the actual resource itself need only maintain as invariant its populated state. Then when you have acquired the resource from the proxy, you always have a guaranteed formed object to work with and the proxy takes care of ensuring it is populated, perhaps in a just-in-time manner.
But this is design nitpicking and probably not very related to the real world. In reality of course a Mesh object will have a vertexCount == 0 state and if this mechanism is used to implement a decision at each use as to whether to load it or not before use, its much the same thing I suppose, except this is unwrapped so has to be performed all over the place.
