.
For starters, let me scream from the highest rooftop that DelayLoad is not an operating system feature. It works on any Win32

A Quick Review
If you're wondering "What's this thing Matt's gone off the deep end over?" a quick recap of DelayLoad is in order. Here's how it works. Normally, when calling an imported function in a DLL, the linker adds information about the imported DLL and function to your executable. Collectively, the information for all the imported functions is known as the imports section.
The Win32 loader scans through the imports section at load time and loads each DLL. For each DLL loaded, the loader iterates through all the imported functions and locates their addresses in the imported DLL. These addresses are written back to the imports section in a location known as the Import Address Table (IAT). A simple way to think of an IAT is as an array of function pointers. When calling an imported function, the call uses one of the function pointers from the IAT.
How does the picture change with DelayLoad? When you specify DelayLoad for a DLL, the linker doesn't emit the usual data it would put in the imports section. Instead, it generates a small stub for each DelayLoad imported function. This stub points to the imported DLL and function name. Upon calling an imported function for the first time, the stub calls LoadLibrary to load the DLL. Next, it calls GetProcAddress to get the address of the called function. Finally, the stub overwrites part of itself so that subsequent calls to the function go directly to the target code.
What I've just described is a slight simplification. In reality, the stub is a small bit of code that calls a routine statically linked into your executable. This routine resides in DELAYIMP.LIB, which must be included in the list of libraries that the linker uses. Also, the stubs and DELAYIMP.LIB code are smart enough to call LoadLibrary only the first time a function in the DLL is used. Subsequent calls to other functions in the same DelayLoad imported DLL don't call LoadLibrary.
All things considered, DelayLoad doesn't add much time or space overhead compared to importing the DLL the usual way. Calling LoadLibrary is only slightly less efficient than letting the Win32 loader load the DLL. Likewise, calling GetProcAddress once for each DelayLoad imported function is only slightly slower than having the Win32 loader locate the imported functions at startup.
However, the benefits of DelayLoad can easily make up for these small speed penalties. For starters, if you never call a function in a DelayLoad imported DLL, the DLL isn't loaded in the first place. This comes in handy more often than you may think. Consider the situation in which you have printing code in your program. If the user doesn't print something during a program session, you've loaded WINSPOOL.DRV for no reason. In this case, using DelayLoad is actually faster since you never loaded and initialized WINSPOOL.DRV.
Another benefit of using DelayLoad is that you avoid calling APIs that are not available on one of your target platforms. For instance, say you want to call AnimateWindow, which is supported in Windows® 98 and Windows 2000, but not Windows 95 or Windows NT 4.0. If you were to call AnimateWindow the usual way, your code wouldn't load on the earlier platforms. However, with DelayLoad you can make a runtime check of which operating system you're on and only call AnimateWindow if it's supported. There's no need for you to muck up your code with calls to LoadLibrary and GetProcAddress.
Using DelayLoad is incredibly easy. Once you know which DLLs you want to use DelayLoad with, simply add /DELAYLOAD:DLLNAME, where DLLNAME is the name of the DLL. You'll also need to add DELAYIMP.LIB to the linker's library list, and you'll still need the original import library, for example, SHELL32.LIB. Putting everything together, to DelayLoad against SHELL32.DLL your linker line would need the following:

SHELL32.LIB /DELAYLOAD:SHELL32.DLL DELAYIMP.LIB

When to Use DelayLoad
When you have a small project, it's easy to come up with a list of DLLs that are good DelayLoad candidates. However, because projects may grow and can involve many developers, it's just as easy to lose track of who uses which DLL. In the past, I've relied on gut instinct and Depends.EXE from the Platform SDK. A DLL from which only a few functions are imported is a good place to start.
However, I wanted a way to automate and simplify the process. Thus was born the DelayLoadProfile program. DelayLoadProfile is a tool that runs your EXE and monitors the DLLs and functions that your EXE calls. After your program terminates, DelayLoadProfile spits out a summary of which DLLs were used and how many calls were made to each DLL. A DLL that's imported, but which had no calls made to it, is a good candidate for DelayLoad importing.
Let me emphasize one point before continuing: DelayLoadProfile works only against your EXE. While it could be extended to recurse into all of your imported DLLs and their dependencies, that would significantly complicate its code. As I'll explain later, DelayLoadProfile just gives you hints about which DLLs you might consider using /DELAYLOAD on. You still have to use that neuron-based processing unit between your ears to make sure it makes sense to do so.

DelayLoadProfile: The Big Picture
The concept behind DelayLoadProfile is simple. Redirecting the function pointers in the EXE's IAT to point to a stub is all that's needed. The stub simply notes that the imported function has been called, then jumps to the address that the Win32 loader originally stored in the IAT. However, the devil is in the details.
First, you must decide where the code will run that locates and modifies the EXE's IAT entries to point to the stubs. Doing the work out-of-process in some sort of control program is one option. This avoids the work involved in getting your code into the target EXE's process. The downside is that it's more work to traverse all the data structures necessary to locate and patch the IAT entries, as well as gather the results later. I'd be swimming in ReadProcessMemory calls.
The other approach is to do the hard work in the same process space as the target EXE. This makes it almost trivial to march through the data structures, build stubs, redirect the IAT entries, and summarize the results at the end. However, doing the work in-process requires that some of the DelayLoadProfile code be loaded into the target EXE's process as it runs. This is the path I took.
Having committed to running my code in-process with the target, the next problem was figuring out how to get my code into the target process. One choice would have been to ask the user to link with the DelayLoadProfile code. Knowing it would require some effort by the target audience, I discarded this option. If a DelayLoadProfile user needed to modify their source, project, or makefile, many would pass. I needed to make DelayLoadProfile a complete no-brainer.
At this point, I had boxed myself into some sort of loader program that would run the target EXE and inject my DelayLoadProfile DLL into it. One technique for DLL injection is to use CreateRemoteThread to start a thread in the target process that calls LoadLibrary on your DLL. I discarded this approach because CreateRemoteThread isn't available on Windows 9x, which I wanted to support.
Longtime MSJ readers may remember a program I wrote more than five years ago called APISPY32. It loads a process and injects a DLL into it for the purposes of logging API calls. That sounds similar to what I needed DelayLoadProfile to do. Alas, when I ran APISPY32 on Windows 2000, it failed to load the DLL. A little digging revealed the source of the problem, and I decided it was time to revamp this code for a whole new generation of programmers.

Into the Trenches
To review quickly, DelayLoadProfile is a two-part system. A loader process runs your program. Early on in your program, the loader process injects a DLL into your program's address space. This DLL scans through your EXE's IAT and redirects the imported functions to point to stubs that the DLL creates. When your program shuts down, the injected DLL scans through the stubs it has created and summarizes how many calls were made to each imported DLL. If you've ever used the APIMON utility from the Platform SDK, you'll recognize the similarities.
The DLL that does all the work of monitoring a program's use of imports is called DelayLoadProfileDLL (see Figure 1). DelayLoadProfileDLL uses the DLL_PROCESS_ATTACH and DLL_PROCESS_DETACH notifications sent to its DllMain procedure to initiate the two primary phases of the DLL's work.
When its DllMain gets the DLL_PROCESS_ATTACH notification, DelayLoadProfileDLL calls PrepareToProfile. Inside PrepareToProfile, the code locates the target EXE's IAT. For each imported DLL it finds, the code determines if it's a DLL that's safe for IAT redirection. It does this by calling the IsModuleOKToHook function. Most of the time, it's OK to redirect the IAT, so PrepareToProfile invokes the RedirectIAT function.
RedirectIAT is where things get dirty, and it really helps if you understand the import-related data structures in WINNT.H. First, the function locates the IAT and the associated Import Names Table. The code then counts how many IAT entries there are by scanning through the IAT, looking for a NULL pointer. With this count, an array of DLPD_IAT_STUB stubs is created, with one stub for each IAT entry.
Finally, it's time for meatball surgery. The code makes yet another pass through the IAT. This time it grabs the address in each IAT entry, stuffs it into a JMP instruction in the stub, and redirects the IAT entry to point to the stub. As the code advances through each subsequent IAT entry, it also advances to the next DLPD_IAT_STUB stub in the allocated array. I'll explain DLPD_IAT_STUB stubs a little later in this column.
Two aspects of redirecting the IAT entries to the allocated stubs are worth mentioning. First, the IAT is often placed in a read-only section of the EXE. Ordinarily, an attempt to modify such an IAT pointer would result in an access violation. Luckily, the VirtualProtect API comes to the rescue and enables you to modify the attributes of a target address, in this case, the IAT. Read-write is the attribute you're looking to modify. When it's finished, the code restores the original memory protection attributes.
The other tricky part of redirecting the IAT occurs when you encounter a data import. Although programmers don't frequently do so, it's relatively easy to import data in addition to code. The Visual C++ runtime library DLL (MSVCRT.DLL) has data exports. Redirecting an IAT entry that refers to data in an imported DLL is almost certainly a recipe for problems.
So how do you determine whether an import is a normal code import or a data import? A commercial product could implement a sophisticated algorithm to determine the import type of an IAT entry. However, I took a shortcut and used IsBadWritePtr. If the IAT points to memory that's writeable, it's probably pointing to data. Likewise, if it points to read-only memory, odds are that it's pointing to code. Is this a perfect test? No, but it's good enough for DelayLoadProfile's needs.
Now let's take a look at the stubs. The DLPD_IAT_STUB structure in DelayLoadProfileDLL.H contains the layout, which is a mixture of code and data. Simplifying this structure, a DLPD_IAT_STUB stub looks like this:

 CALL    DelayLoadProfileDLL_UpdateCount 
 JMP     XXXXXXXX // original IAT address 
 DWORD   count 
 DWORD   pszNameOrOrdinal 

When the EXE calls one of the redirected functions, control goes to the CALL instruction in the stub. The DelayLoadProfileDLL_UpdateCount routine in DelayLoadProfileDLL.CPP simply increments the value of the count field of the stub. After that CALL returns, the JMP instruction transfers control to the original address that was stored in the IAT before I bashed it. Figure 2 shows the big picture after the IAT has been redirected to the stubs.
Assembler junkies might be wondering how the DelayLoadProfileDLL_UpdateCount function knows where the stub's count field is in memory. A quick look at the code shows that DelayLoadProfileDLL_UpdateCount finds the return address pushed on the stack by the CALL instruction. The return address points to the JMP XXXXXXXX instruction following the call. Since the CALL instruction is always five bytes, some pointer arithmetic yields the stub's starting address and easy access to the stub's count field.
I had one problem using the DelayLoadProfileDLL_UpdateCount code that's worth mentioning. Originally, the function didn't have the PUSHAD and POPAD instructions to save and restore all of the regular CPU registers. The code worked fine on many programs, but just blew up on others. Finally, I narrowed it down to programs that imported __CxxFrameHandler and _EH_prolog from MSVCRT.DLL. Both of these APIs expect the EAX register to be set to a given value, and DelayLoadProfileDLL_UpdateCount was trashing EAX.
Since the trashed EAX was the problem, I added PUSHAD and POPAD. Alas, the problem remained. In frustration, I examined the compiler-generated code, and then smacked my forehead. Normally when generating code for a debug build, the Visual C++ 6.0 compiler inserts code in the function prolog to set all local variables to the value 0xCC. This code was trashing EAX before my PUSHAD got a chance to execute. To get around this, I had to remove the /GZ option from the debug build settings for DelayLoadProfileDLL.

Reporting Results
As your process shuts down, the system sends the DLL_ PROCESS_DETACH notification to all loaded DLLs. DelayLoadProfileDLL uses this opportunity to harvest the information collected during the run. In a nutshell, this means scanning through all the stub arrays, counting the number of calls that were made through the stubs, and reporting what it finds.
During the setup phase when DelayLoadProfileDLL was redirecting the IATs, it stashed away the address of the EXE's IAT into a global variable (g_pFirstImportDesc). At shutdown time, ReportProfileResults uses this pointer to walk through the imports section again. For each imported DLL, it retrieves the address of the DLL's first IAT entry. If this is an IAT that I've redirected, the first pointer in the IAT should point to the first of the DLPD_IAT_STUB stubs allocated for that DLL. Of course, the code does some sanity checking to ensure that this is the case. If something doesn't look right, DelayLoadProfileDLL ignores that particular imported DLL.
Generally though, everything looks fine, and the first IAT entry points to my stubs. The code then iterates through all the stubs for the DLL. At each stub, the value of the stub's count field is added to a running total for the DLL. When the iteration completes, ReportProfileResults formats a string with the name of the DLL and how many calls were made through the stubs. The code uses OutputDebugString to broadcast its findings.

Loading and Injection
The program that loads your EXE and injects DelayLoadProfileDLL.DLL is calledyou guessed itDelayLoadProfile.EXE (the source code is available from the MSJ Web site at http://www.microsoft.com/msj). This code mainly drives the CDebugInjector class, which I'll describe shortly. Function main obtains the target EXE's command line and passes it to CDebugInjector::LoadProcess. If the process is created successfully, function main tells CDebugInjector which DLL it wants injected. In this case, it's DelayLoadProfileDLL.DLL, which should be located in the same directory as DelayLoadProfile.EXE.
The last step before letting the target run wild is to call CDebugInjector::SetOutputDebugStringCallback. When DelayLoadProfileDLL reports its results via OutputDebugString, CDebugInjector sees them and passes them to the callback you registered. This callback just printfs the strings to the console. Finally, function main calls CDebugInjector::Run. This call lets the target process begin and, when the time is right, injects the DLL into it.
Figure 3 shows The CDebugInjector class. This is where all the good stuff happens. CDebugInjector::LoadProcess creates the specified process as a debugee process. The ramifications of running as a debugee process have been discussed in many articles and in the MSDN documentation, so I won't go into all the details here.
For the purposes of this column, it's sufficient to say that the debugger process (in this case, DelayLoadProfile) has to enter a loop that calls WaitForDebugEvent and ContinueDebugEvent until the debugee terminates. Every time WaitForDebugEvent returns, something has happened in the debugee. This might be an exception (including breakpoints), a DLL load, a thread creation, or other event. The WaitForDebugEvent documentation covers all the events that might occur. The CDebugInjector::Run method contains the code for this loop.
So how does running the target process as a debugee help you inject a DLL? A debugger process has excellent control over the debugee process's execution. Every time a significant event occurs in the debugee, it is suspended until the debugger calls ContinueDebugEvent. Knowing this, a debugger process can add code to the debugee's address space and temporarily change the debugee's registers so that the added code executes.
In more specific terms, CDebugInjector synthesizes a small code stub that calls LoadLibrary. The DLL name parameter to LoadLibrary points to the name of the DLL to inject. CDebugInjector writes the stub (and the associated DLL name) to the debugee's address space. It then calls SetThreadContext to change the debugee's instruction pointer (EIP) to execute the LoadLibrary stub. All of this dirty work occurs within the CDebugInjector::PlaceInjectionStub method.
Immediately following the LoadLibrary call in the stub is a breakpoint instruction (INT 3). This stops the debugee and gives control back to the debugger process. The debugger then uses SetThreadContext again to restore the instruction pointer and other registers to their original values. Another call to ContinueDebugEvent and the debugee is on its way with the DLL injected, none the wiser that anything has happened.
If you don't think too hard, this injection process doesn't sound too messy. Nonetheless, a few interesting problems crop up that complicate things. For example, when is the proper time to create the stub code and redirect control to it? You can't do this immediately after the CreateProcess call because, among other reasons, the imported DLLs haven't been mapped into memory at this point and the EXE's IAT hasn't been fixed up by the Win32 loader. In other words, it's too early.
The solution I ultimately decided on was to let the debugee run until it encounters its first breakpoint. Then I set a breakpoint of my own at the entry point of the EXE. When this second breakpoint triggers, CDebugInjector knows that DLLs in the target process (including KERNEL32.DLL) have initialized, but no code in the EXE has run. This is the perfect time for injecting DelayLoadProfileDLL.DLL.
Incidentally, where does the first breakpoint come from? By definition, a Win32 process that's being debugged calls DebugBreak (also known as INT 3) very early in its execution. In my ancient APISPY32 code, I used the initial DebugBreak as the occasion to do the injection. Unfortunately in Windows 2000, this DebugBreak occurs before KERNEL32.DLL is initialized. Thus, CDebugInjector sets its own breakpoint to go off when the EXE is about to get control, and thus knows that KERNEL32.DLL has been initialized.
Earlier, I mentioned a breakpoint that occurs after the LoadLibrary call returns. This is a third breakpoint for CDebugInjector to handle. All of the mechanics for handling the different breakpoints can be seen in CDebugInjector::HandleException.
Another interesting problem to address with DLL injection is where to write the LoadLibrary stub. Under Windows NT 4.0 and later you can allocate space in another process with VirtualAllocEx, so I took that route. That leaves out Windows 9x, which doesn't support VirtualAllocEx. For this scenario, I took advantage of a unique property of Windows 9x memory-mapped files. These files are visible in all address spaces, and at the same address. I simply create a small memory-mapped file using the system page file as backing, and blast the LoadLibrary stub into it. The stub is implicitly accessible in the debugee process. For the details, see the code listing for CDebugInjector::GetMemoryForLoadLibraryStub at the link at the top of this article.

Using DelayLoadProfile
DelayLoadProfile is a command-line program that writes its results to standard output. From a command prompt, run DelayLoadProfile, specifying the target program and any arguments it needs, such as:

DelayLoadProfile notepad c:\autoexec.bat 

 [d:\column\col66\debug]delayloadprofile calc
 DelayLoadProfile: SHELL32.dll was called 0 times
 DelayLoadProfile: MSVCRT.dll was called 9 times
 DelayLoadProfile: ADVAPI32.dll was called 0 times
 DelayLoadProfile: GDI32.dll was called 60 times
 DelayLoadProfile: USER32.dll was called 691 times