views:

2492

answers:

9

I'm trying to export classes from a DLL that contain objects such as std::vectors and std::stings - the whole class is declared as dll export through:

    class DLL_EXPORT FontManager
{

The problem is that for members of the complex types I get this warning:

warning C4251: 'FontManager::m__fonts' : class 'std::map<_Kty,_Ty>' needs to have dll-interface to be used by clients of class 'FontManager' with [ _Kty=std::string, _Ty=tFontInfoRef ]

I'm able to remove some of the warnings by putting the following forward class declaration before them even though I'm not changing the type of the member variables themselves:

template class DLL_EXPORT std::allocator<tCharGlyphProviderRef>;
template class DLL_EXPORT std::vector<tCharGlyphProviderRef,std::allocator<tCharGlyphProviderRef> >;
std::vector<tCharGlyphProviderRef> m_glyphProviders;

Looks like the forward declaration "injects" the DLL_EXPORT for when the member is compiled but is it safe? Does it realy change anything when the client compiles this header and uses the std container on his side? Will it make all future uses of such a container DLL_EXPORT (and possibly not inline?)? And does it really solve the problem that the warning tries to warn about?

Is this warning anything I should be worried about or would it be best to disable it in the scope of these constructs? The clients and the dll will always be built using the same set of libraries and compilers and those are header only classes...

I'm using Visual Studio 2003 with the standard STD library.

---- Update ----

I'd like to target you more though as I see the answers are general and here we're talking about std containers and types (such as std::string) - maybe the question really is:

Can we disable the warning for standard containers and types available to both the client and the dll through the same library headers and treat them just as we'd treat an int or any other built-in type? (It does seem to work correctly on my side.) If so would should be the conditions under which we can do this?

Or should maybe using such containers be prohibited or at least ultra care taken to make sure no assignment operators, copy constructors etc will get inlined into the dll client?

In general I'd like to know if you feel designing a dll interface having such objects (and for example using them to return stuff to the client as return value types) is a good idea or not and why - I'd like to have a "high level" interface to this functionality... maybe the best solution is what Neil Butterworth suggested - creating a static library?

+6  A: 

When you may touch a member in your class from the client, you need to provide an DLL-interface. A DLL-interface means, that the compiler creates the function in the DLL itself and makes it importable.

Because the compiler doesn't know which methods are used by the clients of an DLL_EXPORTED class it must enforce that all methods are dll-exported. It must enforce that all members which can be accessed by clients must dll-export their functions too. This happens by the compiler warning you of methods not exported and the linker of the client sending errors.

No every member must be marked with with dll-export, e.g. private members not touchable by clients. Here you can ignore/disable the warnings (beware of compiler generated dtor/ctors).

Otherwise the members must export their methods. Forward declaring them with DLL_EXPORT does not export the methods of these classes. You have to mark the according classes in their compilation-unit as DLL_EXPORT.

What it boils down to ... (for not dll-exportable members)

  1. If you have members which aren't/can't be used by clients, switch of the warning.

  2. If you have members which must be used by clients, create an dll-export wrapper or create indirection methods.

  3. To cut down the count of externally visible members use methods like PIMPL.


template class DLL_EXPORT std::allocator<tCharGlyphProviderRef>;

This does create an instantiation of the template specialization in the current compilation unit. So this creates the methods of std::allocator in the dll and exports the corresponding methods. This does not work for concrete classes as this is a only an instantiation of template classes.

Christopher
Thanks Chris - in this case there's the additional issue that those are standard containers that are header only - so I'm not sure we'd need the functions to be called through the dll on those members as long as those functions are identical in both the client and the dll - long story short mine are NOT exposed to the clients directly so I will disable the warnings :) - but are you sure it would hurt to disable them even if they were exposed given the circumstances?
RnR
Btw - there's one problem with defining that classes are private so they can not be touched - this does not take into account copying/creating/destroying of those objects - so we can say that we can disable the warning only if those operations are not directly possible for the client on our main object. In this case though I'm almost sure we don't have any problem at all as the functions are available to the client just as they are to the dll - through the compilers standard headers - right? :)
RnR
I think template classes may not expose these problems, because the methods get instantiated in the clients implementation modules. This does also work for inline functions. But I don't know if all problems go away with that, because some templates call non-templated code (e.g. std::exception::what).
Christopher
"this does not take into account copying/creating/destroying" yes, as mentioned. But declaring these operations does circumvent the problems.
Christopher
+3  A: 

One alternative that few people seem to consider is not to use a DLL at all but to link statically against a static .LIB library. If you do that, all the issues of exporting/importing go away (though you will still have name-mangling issues if you use different compilers). You do of course lose the features of the DLL architecture, such as run-time loading of functions, but this can be a small price to pay in many cases.

anon
You really can't mix compilers anyway, if your DLL interface uses C++ classes - especially if you're using standard library types.
MSalters
There's always a number of ways to skin a cat and if this is an option I agree it should be taken.
RnR
A: 

In such cases, consider the uses of the pimpl idiom. Hide all the complex types behind a single void*. Compiler typically fails to notice that your members are private and all methods included in the DLL.

MSalters
A: 

If you use a DLL make initialization of all objects at event "DLL PROCESS ATTACH" and export a pointer to its classes/objects.
You may provide specific functions to create and destroy objects and functions to obtain the pointer of the objects created, so you can encapsulate these calls in a wrapper class of access at include file.

lsalamon
+5  A: 

That warning is telling you that users of your DLL will not have access to your container member variables across the DLL boundary. Explicitly exporting them makes them available, but is it a good idea?

In general, I'd avoid exporting std containers from your DLL. If you can absolutely guarantee your DLL will be used with the same runtime and compiler version you'd be safe. You must ensure memory allocated in your DLL is deallocated using the same memory manager. To do otherwise will, at best, assert at runtime.

So, don't expose containers directly across DLL boundaries. If you need to expose container elements, do so via accessor methods. In the case you provided, separate the interface from the implementation and expose the inteface at the DLL level. Your use of std containers is an implementation detail that the client of your DLL shouldn't need to access.

Alternatively, do what Neil suggest and create a static library instead of a DLL. You lose the ability to load the library at runtime, and consumers of your library must relink anytime you change your library. If these are compromises you can live with, a static library would at least get you past this problem. I'll still argue you're exposing implementation details unnecessarily but it might make sense for your particular library.

Aaron Saarela
+1  A: 

Found this article. In short Aaron has the 'real' answer above; Don't expose standard containers across library boundaries.

Chris Huang-Leaver
Thanks Chris - it's really a gread article and basically sums up this thread nicely - looks we ( disable it and assume compile environment compatibility between clients and the dll OR use lib's and not dll's OR don't use templates in exported classes ) AND none of the above seem right :(
RnR
+1  A: 

Though this thread is pretty old, I found a problem recently, which made me think again about having templates in my exported classes:

I wrote a class which had a private member of type std::map. Everything worked quite well untill it got compiled in release mode, Even when used in a build system, which ensures that all compiler settings are the same for all targets. The map was completely hidden and nothing was directly exposed to the clients.

As a result the code was just crashing in release mode. I gues, because different binary std::map instances were created for implementation and client code.

I guess the C++ Standard is not saying anaything about how this shall be handled for exported classes as this is pretty much compiler specific. So I guess the biggest portability rule is to just expose Interfaces and use the PIMPL idiom as much as possible.

Thanks for any enlightenment

TheBigW
A: 

none of the workarounds above are acceptable with MSVC because of the static data members inside template classes like stl containers

each module (dll/exe) gets its own copy of each static definition...wow! this will lead to terrible things if you somehow 'export' such data (as 'pointed' above)...so don't try this at home

see http://support.microsoft.com/kb/172396/en-us

tcris
+1  A: 

There are other issues.

Some STL containers are "safe" to export (such as vector), and some aren't (e.g. map).

Map for instance is unsafe because it (in the MS STL distribution anyway) contains a static member called _Nil, the value of which is compared in iteration to test for the end. Every module compiled with STL has a different value for _Nil, and so a map created in one module will not be iterable from another module (it never detects the end and blows up).

This would apply even if you statically link to a lib, since you can never guarantee what the value of _Nil will be (it's uninitialised).

I believe STLPort doesn't do this.

Adrien
Thanks for the details - personally I always preferred STLPort and now I'll have another reason for it :D
RnR