Then you've got two options.

1. Use C functions only.

2. Use C++, along with some ungodly hacks, and revel in the ensuing maintenance nightmare!

The first goals 3 are achievable, but requirement 4 is a problem. The simplest way of doing this is with pure virtual interfaces (because you only need to export a single C function to create an instance of the main device class). To achieve all 4 goals, you'd need to use dbghelp.dll to unmangle the names returned from the DLL, and then find a way to invoke those methods via member function pointers (to classes you can't use, because you can't link to them). The (now ancient, and not particularly great for new projects) function binding interface from Scott Bilas (http://scottbilas.com/files/2001/gdc_san_jose/fubi_slides_outline.pdf) might provide some pointers, although expect to spend a lot of time digging through the PLE file, and how that relates to x64 (the code is 32bit). It's possible, but really really fugly.

A much simpler (and more efficient solution), is to simply use static libs for the renderer code. Compile D3D and OGL versions of your app, and then use a 3rd 'game-config' app to execute the relevant exe (based on whether the user has selected. D3D or GL). Any other libs you are using that aren't dependent on rendering code, should be fairly easy to move to DLLs that can be reused by both executables, which kinda gives you the same result, but somewhat in reverse (and you'll retain your sanity)

\edit

I've just re-read your question, and I may have misread the bit about runtime switching. Ok, so if runtime switching is not a goal, then yes it's possible via the pimpl idiom (which should bury the platform specific stuff deep within the DLL, and leave your exported classes to expose only themselves [rather than their platform specific members]). Generally though, you'll end up with exactly the same amount of indirection as if you'd used pure virtual interfaces, and the code will be even more annoying to use in practice (pure virtual interfaces are the easiest way to achieve this goal, which will leave you with a runtime switchable DLL as a bonus). There are other ways to ensure ABI compatibility between DLL's, but in this particular case, your code would need to conform to some extremely rigid coding conventions (As an aside, I started putting together into an article for gamedev that explains how to do this, but that's on hold for the next few weeks whilst I move house!). Either don't bother, or go with virtual interfaces, or with pimpl. Those are your best options imho.

I've done the whole renderer-in-a-runtime-swappable-dll-accessed-via-pure-virtual-interface thing in the past.

I can't be sure why you don't want to use interface inheritance, but I can say that if my engine were making so many calls to the renderer that the virtual function call overhead became a performance concern, I'd think I must be doing something very wrong.

So in short, i'm trying to:
1. link my renderer DLL into my main project
2. only link the d3d libs in the renderer DLL (EDIT: oh, and that includes the include/lib/source directories as well)
3. not have to link d3d in the main project
4. not use inheritance for any of the graphics classes, but still be able to use them in the main project (and not use any API specific calls on those classes, mostly just getters)

I hope i explained myself well enough to get some advice. Thanks in advance!

If that is all, than use dx,ogl headers only in renderer definitions (cpp files) and never expose them to declarations of rendered classes. That is all that's needed to do. You will have to write a lot more code to recieve such an isolation with enough level of operatibility though. And when you do, watch out for optimization side, (avoid memory copying and other unneeded stuff).

I thought i'd have to expose all the low level stuff (like vertex/index buffer, textures and such), but if i compose them into higher level concepts (like sprite, model etc), than i end up with just a few interfaces. Is this a good approach?

Whichever approach is going to be easiest for you. There's no right or wrong way really.

While i'm at it, i've got another question. If i use the interfaces to call the underlying concrete objects, that means that all of them have to be heap allocated, unless i use globals inside the dll (and i don't want to if i can help it)?

Also, everything created within the dlls factory method should also be sent to the dll to destroy instead of destroying them manually, right?

You can call delete on the objects, but you WILL need a virtual destructor. I still prefer a destroy method myself, but that's just me.

And lastly, the runtime switching, is that possible if i link the .lib and use the DLL for implementation, or do i need to use LoadLibrary and manually do the switching?

Yes, you will need to use LoadLibrary (or dlopen for posix). *If* you have a pure virtual base class, then the minimum you'd need to faff around with is:

// plugin API

[source]

class Device
{
public:

virtual ~Device() {}

};

extern "C"
{
typedef Device* (*createDeviceFunc)();

}

[/source]

// plugin Implementation

[source]
class MyDevice : public Device
{
public:
MyDevice() {}
}:

extern "C"
{
__declspec(dllexport) Device* createDevice() { return new MyDevice; }

}

[/source]

// in your app

[source]

#include "PluginAPI.h"

createDeviceFunc createDevice = 0;
HMODULE plugin = 0;

Device* device = 0;

void loadPlugin(const char* dllname)

{

plugin = LoadLibraryA( dllname );
if(plugin)
{
createDevice = (createDeviceFunc)GetProcAddressA(plugin, "createDevice");

if(createDevice)
{

device = createDevice();
}
}
}

// then somewhere in your startup code
loadPlugin("myGLRenderer.dll");

[/source]

I'm actually not sure if its even possible to change the DLL on the fly while the exe is running, as the DLL would be in use, right?

You'll need to make sure it's no longer in use, and then it can be unloaded safely.

[source]

void unloadPlugin()

{
delete device; //< this must also delete all buffers, textures, etc
device = 0;
createDevice = 0;
FreeLibrary(plugin);

}

[/source]

And even if that's possible, how would the exe know to reload the DLL?

You'll have to tell it to load the plugin again!

[source]
loadPlugin("myD3D11Renderer.dll");

[/source]

What i wanted to avoid is having to create an interface for each class that i'd put into the graphics dll, but i haven't made a list of what classes from the dll would actually be exposed outward. I thought i'd have to expose all the low level stuff (like vertex/index buffer, textures and such), but if i compose them into higher level concepts (like sprite, model etc), than i end up with just a few interfaces. Is this a good approach?

Yes.  You should keep the renderer interface used by the rest of the engine as minimal and simple as possible.

Also, everything created within the dlls factory method should also be sent to the dll to destroy instead of destroying them manually, right?

No.  Your dll will generally be compiled against the same memory manager as the rest of the engine (unless you do something to prevent it), and your engine is linked against the interface you expose from the dll, so generally, the engine can delete objects it gets from the dll.

That being said, code defined in the dll, including destructors and anything they call, can only be called while the dll is loaded.  So it's a bad idea for the renderer to be handing out objects implemented in the dll.

In my implementation, the engine only handled a single dll specific object - the renderer itself.  All other render related objects handled by the engine were defined outside of the dll.  They were written in terms of the renderer's interface.  All data resources with API specific formats were held within the renderer internally, and referenced externally by ID (so the resource could swap out when the renderer did, without breaking external references).

On a side note, this means that if you ever do decide to make your renderer hot-swappable, it'll need to do a little fancy footwork for the post-swap renderer to acquire all the API specific data resources corresponding to those which had been held by the pre-swap renderer.

I'm actually not sure if its even possible to change the DLL on the fly while the exe is running, as the DLL would be in use, right? And even if that's possible, how would the exe know to reload the DLL?

Yes, you can unload a dll.  It's up to you to make sure you don't use it before loading or after unloading it.

I'm not sure I understand the question - but the exe "knows" because you write it to know.  You unload one dll and load another when you see fit.

Btw, on a completely unrelated note to DLLs, but related to Direct3D itself, is there any "standard" (or a majority opinion) for using COM smart pointers?

If DLL's are involved, I'd avoid smart pointers completely (personally speaking). If you *really* need that kind of thing, and it works for you, then using CComPtr will be better than the alternatives (i.e. boost et al).

...in my mind i was referring to hot swapping the exact same DLL during runtime...

Ah, I get it.

It's been a while, so I don't know for certain, but it seems like the running program shouldn't be holding the dll file open after it loads it.  If not, you should be able to overwrite the file after it's loaded.  Then you could have the engine periodically check its modified date, and initiate swap proceedings when it sees that date change.

With a little attention to data handling and architecture you should be able to update lots of different systems on the fly without restarting the engine.  You could also choose to load dlls only in a development build, and link to static libraries for a release build.