.NET Interop IntPtr vs. ref

Question

Probably a noob question but interop isn't one of my strong points yet.

Aside from limiting the number of overloads is there any reason I should declare my DllImports like:

[DllImport("user32.dll")]
public static extern int SendMessage(IntPtr hWnd, int msg, int wParam, IntPtr lParam);

And use them like this:

IntPtr lParam = Marshal.AllocCoTaskMem(Marshal.SizeOf(formatrange));
Marshal.StructureToPtr(formatrange, lParam, false);

int returnValue = User32.SendMessage(_RichTextBox.Handle, ApiConstants.EM_FORMATRANGE, wParam, lParam);

Marshal.FreeCoTaskMem(lParam);

Rather than creating a targeted overload:

[DllImport("user32.dll")]
public static extern int SendMessage(IntPtr hWnd, int msg, int wParam, ref FORMATRANGE lParam);

And using it like:

FORMATRANGE lParam = new FORMATRANGE();
int returnValue = User32.SendMessage(_RichTextBox.Handle, ApiConstants.EM_FORMATRANGE, wParam, ref lParam);

The by ref overload ends up being easier to use but I'm wondering if there is a drawback that I'm not aware of.

Edit:

Lots of great info so far guys.

@P Daddy: Do you have an example of basing the struct class off an abstract (or any) class? I changed my signature to:

[DllImport("user32.dll", SetLastError = true)]
public static extern int SendMessage(IntPtr hWnd, int msg, int wParam, [In, Out, MarshalAs(UnmanagedType.LPStruct)] CHARFORMAT2 lParam);

Without the In, Out, and MarshalAs the SendMessage (EM_GETCHARFORMAT in my test) fail. The above example works well but if I change it to:

[DllImport("user32.dll", SetLastError = true)]
public static extern int SendMessage(IntPtr hWnd, int msg, int wParam, [In, Out, MarshalAs(UnmanagedType.LPStruct)] NativeStruct lParam);

I get a System.TypeLoadException that says the CHARFORMAT2 format is not valid (I'll try and capture it for here).

The exception:

Could not load type 'CC.Utilities.WindowsApi.CHARFORMAT2' from assembly 'CC.Utilities, Version=1.0.9.1212, Culture=neutral, PublicKeyToken=111aac7a42f7965e' because the format is invalid.

The NativeStruct class:

public class NativeStruct
{
}

I've tried abstract, adding the StructLayout attribute, etc. and I get the same exception.

[StructLayout(LayoutKind.Sequential)]
public class CHARFORMAT2: NativeStruct
{
    ...
}

Edit:

I didn't follow the FAQ and I asked a question that can be discussed but not positively answered. Aside from that there has been lot's of insightful information in this thread. So I'll leave it up to the readers to vote up an answer. First one to over 10 up-votes will be the answer. If no answer meets this in two days (12/17 PST) I'll add my own answer that summarizes all the yummy knowledge in the thread :-)

Edit Again:

I lied, accepting P Daddy's answer because he is the man and has been a great help (he has a cute little monkey too :-P)

Answer 1

I don't see any drawbacks.

By-ref is often enough for simple type and simple structure.

IntPtr should be favored if the structure has a variable-size or if you want to do custom processing.

Answer 2

I've had some fun cases where a parameter is something like ref Guid parent and the corresponding documentation says "Pointer to a Guid specifying the parent. Pass a null pointer to use [insert some system defined item]."

If null (or IntPtr.Zero for IntPtr parameters) really is an invalid parameter, then you're fine using a ref parameter - maybe even better off since it's extra clear exactly what you need to pass.

Edit: If null is a valid parameter, you can pass ClassType instead of ref StructType. Objects of a reference type (class) are passed as a pointer, and they allow null.

Answer 3

Using ref is simpler and less error-prone than manipulating pointers manually, so I see no good reason for not using it... Another benefit of using ref is that you don't have to worry about freeing unmanaged allocated memory

Answer 4

If the struct is marshalable without custom processing, I greatly prefer the latter approach, where you declare the p/invoke function as taking a ref (pointer to) your type. Alternatively, you can declare your types as classes instead of structs, and then you can pass null, as well.

[StructLayout(LayoutKind.Sequential)]
struct NativeType{
    ...
}

[DllImport("...")]
static extern bool NativeFunction(ref NativeType foo);

// can't pass null to NativeFunction
// unless you also include an overload that takes IntPtr

[DllImport("...")]
static extern bool NativeFunction(IntPtr foo);

// but declaring NativeType as a class works, too

[StructLayout(LayoutKind.Sequential)]
class NativeType2{
    ...
}

[DllImport("...")]
static extern bool NativeFunction(NativeType2 foo);

// and now you can pass null

<pedantry>

By the way, in your example passing a pointer as an IntPtr, you've used the wrong Alloc. SendMessage is not a COM function, so you shouldn't be using the COM allocator. Use Marshal.AllocHGlobal and Marshal.FreeHGlobal. They're poorly named; the names only make sense if you've done Windows API programming, and maybe not even then. AllocHGlobal calls GlobalAlloc in kernel32.dll, which returns an HGLOBAL. This used to be different from an HLOCAL, returned by LocalAlloc back in the 16-bit days, but in 32-bit Windows they are the same.

The use of the term HGLOBAL to refer to a block of (native) user-space memory just kind of stuck, I guess, and the people designing the Marshal class must not have taken the time to think about how unintuitive that would be for most .NET developers. On the other hand, most .NET developers don't need to allocate unmanaged memory, so....

</pedantry>

---

Edit

You mention you're getting a TypeLoadException when using a class instead of a struct, and ask for a sample. I did up a quick test using CHARFORMAT2, since it looks like that's what you're trying to use.

First the ABC¹:

[StructLayout(LayoutKind.Sequential)]
abstract class NativeStruct{} // simple enough

The StructLayout attribute is required, or you will get a TypeLoadException.

Now the CHARFORMAT2 class:

[StructLayout(LayoutKind.Sequential, Pack=4, CharSet=CharSet.Auto)]
class CHARFORMAT2 : NativeStruct{
    public DWORD    cbSize = (DWORD)Marshal.SizeOf(typeof(CHARFORMAT2));
    public CFM      dwMask;
    public CFE      dwEffects;
    public int      yHeight;
    public int      yOffset;
    public COLORREF crTextColor;
    public byte     bCharSet;
    public byte     bPitchAndFamily;
    [MarshalAs(UnmanagedType.ByValTStr, SizeConst=32)]
    public string   szFaceName;
    public WORD     wWeight;
    public short    sSpacing;
    public COLORREF crBackColor;
    public LCID     lcid;
    public DWORD    dwReserved;
    public short    sStyle;
    public WORD     wKerning;
    public byte     bUnderlineType;
    public byte     bAnimation;
    public byte     bRevAuthor;
    public byte     bReserved1;
}

I've used using statements to alias System.UInt32 as DWORD, LCID, and COLORREF, and alias System.UInt16 as WORD. I try to keep my P/Invoke definitions as true to SDK spec as I can. CFM and CFE are enums that contain the flag values for these fields. I've left their definitions out for brevity, but can add them in if needed.

I've declared SendMessage as:

[DllImport("user32.dll", CharSet=CharSet.Auto)]
static extern IntPtr SendMessage(
    HWND hWnd, MSG msg, WPARAM wParam, [In, Out] NativeStruct lParam);

HWND is an alias for System.IntPtr, MSG is System.UInt32, and WPARAM is System.UIntPtr.

[In, Out] attribute on lParam is required for this to work, otherwise, it doesn't seem to get marshaled both directions (before and after call to native code).

I call it with:

CHARFORMAT2 cf = new CHARFORMAT2();
SendMessage(rtfControl.Handle, (MSG)EM.GETCHARFORMAT, (WPARAM)SCF.DEFAULT, cf);

EM and SCF are enums I've, again, left out for (relative) brevity.

I check success with:

Console.WriteLine(cf.szFaceName);

and I get:

Microsoft Sans Serif

Works like a charm!

---

Um, or not, depending on how much sleep you've had and how many things you're trying to do at once, I suppose.

This would work if CHARFORMAT2 were a blittable type. (A blittable type is a type that has the same representation in managed memory as in unmanaged memory.) For instance, the MINMAXINFO type does work as described.

[StructLayout(LayoutKind.Sequential)]
class MINMAXINFO : NativeStruct{
    public Point ptReserved;
    public Point ptMaxSize;
    public Point ptMaxPosition;
    public Point ptMinTrackSize;
    public Point ptMaxTrackSize;
}

This is because blittable types are not really marshaled. They're just pinned in memory—this keeps the GC from moving them—and the address of their location in managed memory is passed to the native function.

Non-blittable types have to be marshaled. The CLR allocates unmanaged memory and copies the data between the managed object and its unmanaged representation, making the necessary conversions between formats as it goes.

The CHARFORMAT2 structure is non-blittable because of the string member. The CLR can't just pass a pointer to a .NET string object where a fixed-length character array is expected to be. So the CHARFORMAT2 structure must be marshaled.

As it would appear, for correct marshaling to occur, the interop function must be declared with the type to be marshaled. In other words, given the above definition, the CLR must be making some sort of determination based on the static type of NativeStruct. I would guess that it's correctly detecting that the object needs to be marshaled, but then only "marshaling" a zero-byte object, the size of NativeStruct itself.

So in order to get your code working for CHARFORMAT2 (and any other non-blittable types you might use), you'll have to go back to declaring SendMessage as taking a CHARFORMAT2 object. Sorry I led you astray on this one.

---

Captcha for the previous edit:

the whippet

Yeah, whip it good!

---

Cory,

This is off topic, but I notice a potential problem for you in the app it looks like you're making.

The rich textbox control uses standard GDI text-measuring and text-drawing functions. Why is this a problem? Because, despite claims that a TrueType font looks the same on screen as on paper, GDI does not accurately place characters. The problem is rounding.

GDI uses all-integer routines to measure text and place characters. The width of each character (and height of each line, for that matter) is rounded to the nearest whole number of pixels, with no error correction.

The error can easily be seen in your test app. Set the font to Courier New at 12 points. This fixed-width font should space characters exactly 10 per inch, or 0.1 inches per character. This should mean that, given your starting line width of 5.5 inches, you should be able to fit 55 characters on the first line before wrap occurs.

ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz123

But if you try, you'll see that wrap occurs after only 54 characters. What's more the 54^th character and part of the 53^rd overhang the apparent margin shown on the ruler bar.

This assumes you have your settings at standard 96 DPI (normal fonts). If you use 120 DPI (large fonts), you won't see this problem, although it appears that you size your control incorrectly in this case. You also won't likely see this on the printed page.

What's going on here? The problem is that 0.1 inches (the width of one character) is 9.6 pixels (again, using 96 DPI). GDI doesn't space characters using floating point numbers, so it rounds this up to 10 pixels. So 55 characters takes up 55 * 10 = 550 pixels / 96 DPI = 5.7291666... inches, whereas what we were expecting was 5.5 inches.

While this will probably be less noticeable in the normal use case for a word processor program, there is a likelihood of instances where word wrap occurs at different places on screen versus on page, or that things don't line up the same once printed as they did on screen. This could turn out to be a problem for you if this is a commercial application you're working on.

Unfortunately, the fix for this problem is not easy. It means you'll have to dispense with the rich textbox control, which means a huge hassle of implementing yourself everything it does for you, which is quite a lot. It also means that the text drawing code you'll have to implement becomes fairly complicated. I've got code that does it, but it's too complex to post here. You might, however, find this example or this one helpful.

Good luck!

---

^{1 Abstract Base Class}

Answer 5

No, you cannot overload SendMessage and make the wparam argument an int. That will make your program fail on a 64-bit version of the operating system. It has to be a pointer, either IntPtr, a blittable reference or an out or ref value type. Overloading the out/ref type is otherwise fine.

---

EDIT: As the OP pointed out, this is not actually a problem. The 64-bit function calling convention passes the first 4 arguments through registers, not the stack. There is thus no danger of stack mis-alignment for the wparam and lparam arguments.