Win32 DLLs in D

DLLs (Dynamic Link Libraries) are one of the foundations of system programming for Windows. The D programming language enables the creation of several different types of DLLs.

For background information on what DLLs are and how they work Chapter 11 of Jeffrey Richter's book Advanced Windows is indispensible.

This guide will show how to create DLLs of various types with D.

Compiling a DLL

Use the -shared switch to tell the compiler that the generated code is to be put into a DLL. Code compiled for an EXE file will use the optimization assumption that _tls_index==0. Such code in a DLL will crash.

DLLs with a C Interface

A DLL presenting a C interface can connect to any other code in a language that supports calling C functions in a DLL.

DLLs can be created in D in roughly the same way as in C. A DllMain() is required, looking like:

import std.c.windows.windows;
import core.sys.windows.dll;

__gshared HINSTANCE g_hInst;

extern (Windows)
BOOL DllMain(HINSTANCE hInstance, ULONG ulReason, LPVOID pvReserved)
{
    switch (ulReason)
    {
	case DLL_PROCESS_ATTACH:
	    g_hInst = hInstance;
	    dll_process_attach( hInstance, true );
	    break;

	case DLL_PROCESS_DETACH:
	    dll_process_detach( hInstance, true );
	    break;

	case DLL_THREAD_ATTACH:
	    dll_thread_attach( true, true );
	    break;

	case DLL_THREAD_DETACH:
	    dll_thread_detach( true, true );
	    break;
    }
    return true;
}

Notes:

• DllMain simply forwards to the appropriate helper functions. These setup the runtime, create thread objects for interaction with the garbage collector and initialize thread local storage data.
• The DLL does not share its runtime or memory with other DLLs.
• The first boolean argument to the dll-helper functions specify whether all threads should be controlled by the garbage collector. You might need more control over this behaviour if there are threads in the process that must not be suspended. In this case pass false to disable the automatic handling of all threads.
• The presence of DllMain() is recognized by the compiler causing it to emit a reference to __acrtused_dll and the phobos.lib runtime library.

Link with a .def (Module Definition File) along the lines of:

LIBRARY         MYDLL
DESCRIPTION     'My DLL written in D'

EXETYPE		NT
CODE            PRELOAD DISCARDABLE
DATA            WRITE

EXPORTS
		DllGetClassObject       @2
		DllCanUnloadNow         @3
		DllRegisterServer       @4
		DllUnregisterServer     @5

The functions in the EXPORTS list are for illustration. Replace them with the actual exported functions from MYDLL. Alternatively, use implib. Here's an example of a simple DLL with a function print() which prints a string:

mydll.d:

module mydll;
import std.c.stdio;
export void dllprint() { printf("hello dll world\n"); }

Note: We use printfs in these examples instead of writefln to make the examples as simple as possible.

mydll.def:

LIBRARY "mydll.dll"
EXETYPE NT
SUBSYSTEM WINDOWS
CODE SHARED EXECUTE
DATA WRITE

Put the code above that contains DllMain() into a file dll.d. Compile and link the dll with the following command:

C:>dmd -ofmydll.dll -L/IMPLIB mydll.d dll.d mydll.def
C:>

which will create mydll.dll and mydll.lib. Now for a program, test.d, which will use the dll:

test.d:

import mydll;

int main()
{
   mydll.dllprint();
   return 0;
}

Create an interface file mydll.di that doesn't have the function bodies.

mydll.di:

export void dllprint();

Compile and link with the command:

C:>dmd test.d mydll.lib
C:>

and run:

C:>test
hello dll world
C:>

Memory Allocation

D DLLs use garbage collected memory management. The question is what happens when pointers to allocated data cross DLL boundaries? If the DLL presents a C interface, one would assume the reason for that is to connect with code written in other languages. Those other languages will not know anything about D's memory management. Thus, the C interface will have to shield the DLL's callers from needing to know anything about it.

There are many approaches to solving this problem:

• Do not return pointers to D gc allocated memory to the caller of the DLL. Instead, have the caller allocate a buffer, and have the DLL fill in that buffer.
• Retain a pointer to the data within the D DLL so the GC will not free it. Establish a protocol where the caller informs the D DLL when it is safe to free the data.
• Notify the GC about external references to a memory block by calling GC.addRange.
• Use operating system primitives like VirtualAlloc() to allocate memory to be transferred between DLLs.
• Use std.c.stdlib.malloc() (or another non-gc allocator) when allocating data to be returned to the caller. Export a function that will be used by the caller to free the data.

COM Programming

Many Windows API interfaces are in terms of COM (Common Object Model) objects (also called OLE or ActiveX objects). A COM object is an object who's first field is a pointer to a vtbl[], and the first 3 entries in that vtbl[] are for QueryInterface(), AddRef(), and Release().

For understanding COM, Kraig Brockshmidt's Inside OLE is an indispensible resource.

COM objects are analogous to D interfaces. Any COM object can be expressed as a D interface, and every D object with an interface X can be exposed as a COM object X. This means that D is compatible with COM objects implemented in other languages.

While not strictly necessary, the Phobos library provides an Object useful as a super class for all D COM objects, called ComObject. ComObject provides a default implementation for QueryInterface(), AddRef(), and Release().

Windows COM objects use the Windows calling convention, which is not the default for D, so COM functions need to have the attribute extern (Windows).

So, to write a COM object:

import std.c.windows.com;

class MyCOMobject : ComObject
{
    extern (Windows):
	...
}

The sample code includes an example COM client program and server DLL.

D code calling D code in DLLs

Having DLLs in D be able to talk to each other as if they were statically linked together is, of course, very desirable as code between applications can be shared, and different DLLs can be independently developed.

The underlying difficulty is what to do about garbage collection (gc). Each EXE and DLL will have their own gc instance. While these gc's can coexist without stepping on each other, it's redundant and inefficient to have multiple gc's running. The idea explored here is to pick one gc and have the DLLs redirect their gc's to use that one. The one gc used here will be the one in the EXE file, although it's also possible to make a separate DLL just for the gc.

The example will show both how to statically load a DLL, and to dynamically load/unload it.

Starting with the code for the DLL, mydll.d:

/*
 * MyDll demonstration of how to write D DLLs.
 */

import core.runtime;
import std.c.stdio;
import std.c.stdlib;
import std.string;
import std.c.windows.windows;

HINSTANCE   g_hInst;

extern (C)
{
    void  gc_setProxy(void* p);
    void  gc_clrProxy();
}

extern (Windows)
    BOOL DllMain(HINSTANCE hInstance, ULONG ulReason, LPVOID pvReserved)
{
    switch (ulReason)
    {
        case DLL_PROCESS_ATTACH:
	    printf("DLL_PROCESS_ATTACH\n");
	    Runtime.initialize();
	    break;

        case DLL_PROCESS_DETACH:
	    printf("DLL_PROCESS_DETACH\n");
	    Runtime.terminate();
	    break;

        case DLL_THREAD_ATTACH:
	    printf("DLL_THREAD_ATTACH\n");
	    return false;

        case DLL_THREAD_DETACH:
	    printf("DLL_THREAD_DETACH\n");
	    return false;
    }
    g_hInst = hInstance;
    return true;
}

export void MyDLL_Initialize(void* gc)
{
    printf("MyDLL_Initialize()\n");
    gc_setProxy(gc);
}

export void MyDLL_Terminate()
{
    printf("MyDLL_Terminate()\n");
    gc_clrProxy();
}

static this()
{
    printf("static this for mydll\n");
}

static ~this()
{
    printf("static ~this for mydll\n");
}

/* --------------------------------------------------------- */

class MyClass
{
    char[] concat(char[] a, char[] b)
    {
	return a ~ " " ~ b;
    }

    void free(char[] s)
    {
	delete s;
    }
}

export MyClass getMyClass()
{
    return new MyClass();
}