Ticklish Techs

This looks like such a innocent method but it lead to big trouble in one of my projects. But lets start with when someone would use this method that is declared as a static method in the Assembly class. In the MSDN you can read:

Returns the Assembly of the method that invoked the currently executing method.

In our project we had an assembly with a lot of helper methods. On of these gets resources from the calling assembly. In various places of our code we called this method to get icons or other resources. This method used exactly this GetCallingAssembly() method to figure out what assembly to look for resources.

That worked pretty good in debug mode but exceptions were thrown in release mode. We could not understand what is going on. It became even worse: when we build a release version and tried to debug that version (using Visual Studio Debugger) in worked again. It looked like a heisenbug.

It took us some time to figure out what is also written in MSDN:

If the method that calls the GetCallingAssembly method is expanded inline by the compiler (that is, if the compiler inserts the function body into the emitted Microsoft intermediate language (MSIL), rather than emitting a function call), then the assembly returned by the GetCallingAssembly method is the assembly containing the inline code. This might be different from the assembly that contains the original method. To ensure that a method that calls the GetCallingAssembly method is not inlined by the compiler, you can apply the MethodImplAttribute attribute with MethodImplOptions.NoInlining.

The JIT compiler moves code around to optimize for performance. Small methods (up to about 56 Byte IL-Code if I remember it right) can be inlined where the method call was before. But the compiler does this only in release, not in debug mode. Also when attaching the debugger to our release build the JIT compiler stopped inlining to enable debugging and our bug was gone.

After understanding this, the fix is easy. Just don’t allow the compiler to inline that particular method that calls Assembly.GetCallingAssembly(). Then the method stays in the assembly where the source code is written and everything will be fine.

[MethodImplAttribute(MethodImplOptions.NoInlining)]
public void SomeFunction(int i)
{
    // ...
    var a = Assembly.GetCallingAssembly();
    // ...
}

This attribute does the trick and I recommend to use it on all methods that call GetCallingAssembly() and can be called form another assembly and need the real calling assembly.

WPF comes with Dependency Properties and everybody using WPF will know about these new kind of properties. When you define your own Dependency Properties in your own class you can pretty easy add a property change event handler:

public int MyProperty
{
    get { return (int)GetValue(MyPropertyProperty); }
    set { SetValue(MyPropertyProperty, value); }
}
public static readonly DependencyProperty MyPropertyProperty =
    DependencyProperty.Register("MyProperty",
                        typeof(int),
                        typeof(MainWindow),
                        new UIPropertyMetadata(0,
                            MyPropertyvalueChangeCallback));

private static void MyPropertyvalueChangeCallback
                        (DependencyObject d,
                         DependencyPropertyChangedEventArgs e)
{
}

But how to add such a event handler to an already existing Dependency Property or somewhere else then in the defining class? E.g. to an property of a WPF-Control that was not build by you. Take a standard WPF-TextBox; both the Text and the FontSize properties are Dependency Properties but the TextBox-class only provides a change event for the Text-property. Nevertheless you can get a change event for any Dependency Property:

DependencyPropertyDescriptor dpd =
    DependencyPropertyDescriptor.FromProperty
        (Control.FontSizeProperty, typeof (TextBox));
dpd.AddValueChanged(someTextBox, SomeTextBoxFontSizeChanged);

Every time the FontSizeProperty on the instance someTextBox changes the given method is called. It’s that easy and you can implement this code everywhere not only within the class that defines the property.

Some time ago we blogged about a nifty little .net assembly that enables you to access the basic Windows shell operations in shell32.dll . You may want to do so especially in Win7 because your file operations integrate with the fancy Win7 dialogs, the flashing progress bar etc.

The problem we blogged about in the former post was a missing "Pack"-instruction in the marshalling-description of a structure. I guess we didn’t test it with WinXP64… but some days ago we tests it with Win7/64 and it failed. We figured out that we introduced a new bug that occurred on 64 bit machines only. After some more testing we found out that:

• The original code (without Pack-value) crashes on 32 bit machines (as described in the former post).
• The original code (without Pack-value) works fine on 64 bit machines (but we don’t think the original developer had 64 bits 7 years ago ).
• Our fixed code (with Pack=2) works fine on 32 bit machines.
• Our fixed code (with Pack=2) crashes on 64 bit machines.

It seems that the "SHFILEOPSTRUCT" structure in shell32.dll has different layouts in Win7/64 and Win7/32. I guess the guys at MS aim for performance and tell the compiler to optimize. The compiler optimizes by assuming the optimal Pack-value, which usually is the byte-width of the processor; 4 for 32bit, 8 for 64bit.

What effect does the Pack-value have at all?

The "Pack"-value controls the alignment of the starting addresses of each single member of a structure (often called "offset").

Let’s say you have this struct:

struct MammaMia
{
    public int16 sweet16;
    public byte bite;
    public string tanga;
}

Usually you don’t care for the exact way your data is stored in memory, but when you pass it from .NET to the Win-API you have to make sure Win-API receives and returns exactly the right format. This process is called "Marshalling".

One crucial instruction for doing so is the StructLayout attribute.

[StructLayout(LayoutKind.Sequential, Pack = 1)]
struct MammaMia
{
    public int16 sweet16;
    public byte bite;
    public string tanga;
}

(Note: There are a lot more options for the StructLayout attribute and a lot more attributes that help you at marshalling.)

LayoutKind.Sequential means that your data is represented in memory in the same order as you declared it: First "sweet16", than "bite" and then "tanga".

Pack=1 tells .NET that every byte can be the starting point of the next member. So .NET produces this structure in memory:

Now Pack=2 makes sure that only every 2nd byte can be the starting point of the next member. That leads – of course – to unused bytes in memory:

The higher the Pack-value, the more unused bytes you get:

Is this a waste of memory? Sure. But processors can access those addresses a bit faster than other addresses. So it’s an performance/memory tradeoff.

After some more testing, I figured out that the correct Pack-value for Win7/64 is 8, while the correct Pack-value for Win7/32 is 2 (see former post). Why 2? I’d expect 4 here. I don’t know and Microsoft wouldn’t provide the source code I guess.
Maybe in former Windows version they preferred a middle-course between performance and saving memory.

The simple solution

Now it was obvious how to make die ShellLib-assembly work with 32 and 64bit: Check what machine we run on and use a different marshalling. Since I had to declare different structs here, the major code is copying data back and forth, not very interesting.

The only interesting part here is: How do you check on what kind of machine you run? I was looking for some property at System.Environment, but didn’t find anything. Then Andreas pointed me to this post and the solution is too simple to cross one’s mind:

“Just do a IntPtr.Size (static method) and since IntPtr is designed to be an integer whose size is platform specific, it will be 8 if you are on a 64-bit machine and 4 if you are on a 32-bit machine.“

You can find more information and our fixed code soon at the codeproject.com .

Epilog

You can also do the marshalling all by yourself, starting with memory allocation and ending with freeing memory, just like in the old times. But this is another story and will be told later.

Maybe.

When building applications that mix WPF and Winforms windows (e.g. because you are using a Winforms dialog within a WPF application) you will sometimes need a System.Windows.Forms.IWin32Window interface from an WPF System.Windows.Window. This will be useful when using the WinForms OpenFileDialog and wanting to provide a parent window to the ShowDialog() method that only accepts IWin32Window as parent.

The IWin32Window interface declares only one read-only property named Handle that must be provided by the WPF Window.

Implementing the Interface

It is very simple to just add this interface to an existing WPF Window and implementing the single property:

public partial class Window1 : Window, IWin32Window
{
   public IntPtr Handle
   {
      get
      {
         var interopHelper = new WindowInteropHelper(this);
         return interopHelper.Handle;
      }
   }
}

Using a wrapper class

It does not look as a very good solution to add such an interface to each and every Window in you whole application only to be able to be the parent of a dialog. So instead of implementing the interface directly you could build a wrapper class that implements this interface and use this wrapper when opening a dialog:

public class Wpf32Window : IWin32Window
{
    public IntPtr Handle { get; private set; }

    public Wpf32Window(Window wpfWindow)
    {
        Handle = new WindowInteropHelper(wpfWindow).Handle;
    }
}

Opening the dialog will now look like:

d.ShowDialog( new Wpf32Window(this) );

The grand solution to this problem

I like the wrapper solution more than implementing an interface to each and every window but I do not like the changed ShowDialog() call. It just does not look natural any more. To solve this I created an extension Method to be used with the wrapper class:

public static class CommonDialogExtensions
{
    public static DialogResult ShowDialog
                               (this CommonDialog dialog,
                                     Window parent)
    {
        return dialog.ShowDialog(new Wpf32Window(parent));
    }
}

Now the call to open a WinForms dialog with an WPF parent looks as it should look like:

d.ShowDialog(this);

In fact one OpenFileDialog is enough but the .net framework provides two of them:

1. System.Windows.Forms.OpenFileDialog in System.Windows.Forms.dll
2. Microsoft.Win32.OpenFileDialog in PresentationFramework.dll (that is WPF)

With such a setup one would use the first for WinForms applications and the second for WPF apps. In fact if you create a WPF application and just type in OpenFileDialog you will get number 2.

But the 2nd one (even if it is part of the newer part of the framework) is an old dialog where the 1st will use the new and shiny Windows 7 dialog on Win7. Just look at the following screenshots.

Only the first class uses the Windows 7 dialog and comes with proper support for Libraries etc. Even if the second shows the Libraries in its left part it does not fully support this new Windows 7 feature.

I have no idea why the WPF class uses the old windows API but it does. For best compatibility with all Windows versions (including 7) you should always use the Winforms Dialog even when building WPF apps! It does not hurt to reference the System.Windows.Forms.dll (it is installed an every machine running the .net framework) and you can use that dialog from WPF without any trouble.

You should only know about one little disadvantage when using the WinForms dialog. It only accepts WinForms as a parent window no WPF windows. To overcome this limitation please read my other article about the IWin32Window interface and how to implement / provide it in WPF at http://www.ticklishtechs.net/2009/12/22/how-to-get-a-iwin32window-from-a-wpf-window/

Here is a nice little challenge, Benjamin came up with some months ago: “From a list of numbers find the smallest and the largest one – but do not use more than 1,5 comparisons per number.”

Dear reader: Try it!

The solution is simple, straight forward and very nice. And I didn’t find it.

I found something else. My approach was to look at the largest-number-so-far (max) and the smallest-number-so-far (min) as the boundaries of a region. A number outside this region must be larger than max or smaller than min and causes that boundary to change.

To check if a number (a) was outside that region a little computation came to my mind:

check = (max - a)*(min - a);

You would expect n to be smaller than max, so (max-n) should be positive.

You would expect n to be larger than min, so (n-min) should be positive, too.

Two positive numbers multiplied result in a positive number again.

So, if n was outside the boundaries, check would become negative. I only had to check which boundary was exceeded and that’s it:

if (check < 0)
    if (a > max)
        max = a;
    else
        min = a;

Using this approach the number of comparisons  converges to 1 when the length of the list grows. So I found a solution that is way better than 1,5 comparisons per number, didn’t I?

Boooooh

No, I didn’t. I cheated. In fact  (max-a) and (min-a) are comparisons, they just don’t use > or <.

(Actually it’s the other way around: To compute < or > most processors do a subtraction and compare the result to 0.)

So – if you count the subtractions as well you get 3+ comparisons per number…

The intended solution to the challenge (and the code you should provide in your exam) is:

a = list[count++];
b = list[count++];

if (a<b)
{
    if (a<min) min = a;
    if (b>max) max = b;
}
else
{
    if (b < min) min = b;
    if (a > max) max = a;
}

So – what’s the point of this post?

The point is: My silly approach can be faster than the standard solution. Depending on the type and range of the numbers in the list calculating and probing check needs less time than the comparisons in the standard solution.

It may not be much, but sometimes small advantages matter.  For example: For an integer-list of length 10^8 with values from -10.000 to 10.000 my approach is 0.02 seconds faster (on my current laptop). And has less code.

So if you feel like using it – I won’t charge you.

Imaging a small C# winform app with just one ListBox and the following code:

private void Form1_Load(object sender, EventArgs e)
{
    for (int i = 0; i < 5; i++)
    {
        Thread t = new Thread(() => AddToListbox(i));
        t.Start();
    }
}

private void AddToListbox(int i)
{
    if (this.InvokeRequired)
        this.Invoke(new Action<int>(AddToListbox), i);
    else
        this.listBox1.Items.Add(i);
}

A simple loop iterating over the numbers from 0 to 4 and adding these values asynchronous to a ListBox. What do you expect? I expected the numbers from 0 to 4 shown in the ListBox but in any random order since I do not control the threads in any way.

I didn’t expect any number to appear multiple times and I’m totally surprised to  see the number 5!

But ReSharper gave me a hint that I often saw but never understood:

So what’s going on?

I use the syntax of an Lambda expression instead of a regular function call (e.g. using the ThreadStart class and a delegate). This Lambda expression is not evaluated until the thread uses it. And by this time, the loop can be in its next iteration. If the loop is already finished i will be 5.

That is exactly what R# tries to tell me: “Hey, you are accessing here a variable but change it later on. Maybe that is not a good idea.”. – It isn’t.

The solution

Just make a copy of i before passing it into the expression. This copy must be a private copy that will not be changed later. The easiest way to do so is declaring a variable inside the body of the loop. In every iteration of the loop a new integer will be created on the stack and the Lambda expression will access this one.

for (int i = 0; i < 5; i++)
{
    int copy = i;
    Thread t = new Thread(() => AddToListbox(copy));
    t.Start();
}

For those who look for a quick solution:

Try switching off the win32manifest-switch in Visual Studio.

Here is the long story.

Hi! As you may already know we are working on the best .NET – ARIS interface there is so far. It’s an interface that makes heavy use of an old C-Api, using interop and p/invoke.

In Mid-December Benjamin found a very strange behavior that almost spoiled my Christmas. We had just release Version 1.1 when he called me: “Arisan crashes with Vista64.” I was shocked. “And with Vista32 as well”. I felt annihilated. We had done so much testing with various versions of XP, Vista and it worked fine with any of those OSes.

After some testing I saw that ARIS didn’t deliver any pointers at all. That’s rather bad for an interface to C that uses pointer in every call.  After one more day we had more evidence: We had switched to .NET3.5 and VS2008 some month ago and our old versions, that were compiled using the .NET2.0-compiler, still worked.

I was so terrified. The 3.5-framework should be similar to the 2.0-framework, just some (not so) little enhancements. A bug in the 3.5-compiler?

To provide a quick solution we re-wrote parts of the fancy 3.5-syntax, to make our newest enhancement compile with the 2.0-compiler again. That worked (at least we thought so) and we released version 1.1.1.

Then a very frustrating time began. We read articles over and over and quickly found this very interesting post. The problems described there where exactly what we saw.  So we followed the track of the nxcompat-flag and DEP, but it didn’t lead us anywhere.

The C-Api we interface is well documented, but it calls hundreds of other components and one of them crashed deep inside an undiscovered country.

We read pages of differences between 2.0- and 3.5-compiler, new compiler-flags (should have paid more attention here), any documentation we could find, but nothing seemed to help.

We had a feeling it could have something to do with Vista UAC, but we didn’t find any matching explanations to our problem. Being really desperate we started this thread at the MS forums. As you can read there, no one could provide a solution that worked.

I was really sad, because the last thing I wanted to do was to switch the entire development of arisan back to VS2005. I got prepared for a very sad life from now on.

But five days ago my life turned to happiness again. One more time I started of by googling words like “.NET compiler 2.0 3.5 differences” and one more time I read the new features of the 3.5-compiler. This time I obviously paid more attention to the compiler-flags and read about the /win32manifest and /nowin32manifest – switches. I didn’t have much hope left, but tried to set the /nowin32manifest switch.

It worked! Everything worked! The C-Api delivered pointers again. I tried switching the win32manifest on and off and a day later Benjamin confirmed my results. We had found it. But after a long time of suffering (for me it felt like half a year) we still were skeptical and thought we needed a deeper understanding of the problem.

The win32manifest and UAC

Finally we had something to look for and quickly found valuable information on the great I’m just sayin’ blog. The documentation of the switches helped as well, so here is a summary:

The win32manifest is not the .NET-manifest.

The win32manifest is part of the UAC (user access control). UAC was introduced with Vista as an answer to more and more security threads. One of its ideas is to keep malicious software from secretly starting to work. If you work with Vista you may have noticed a lot of dialogs that ask you for permission to run certain programs. THAT is UAC.

Now the win32manifest inside an application tells the OS what level of permission it needs to run. There are levels like “asInvoker” (default), “highestAvailable” or “requireAdministrator”. In case the OS (and the current user) can provide the requested permission, the program runs as a trusted process (according to the granted access level). Of course, 90+% (just my estimation) of the programs on you Vista-machine don’t have a win32mianifest,  because it’s a new feature and older applications just don’t have it. (And I think a majority of newer applications won’t have it either). In order to be compatible with no-win32manifest applications on one hand and providing security on the other hand, Vista can do a miracle called “virtualization”. It’s a bit like running a process in a sandbox: Vista provides anything the process may need to run: disk space, memory, a registry etc. So the process feels cosy.

But anything provided this way is separated from the main system. The process uses a copy of the registry, writes to a special parts of the HDD drives and accesses the memory in a controlled way (not sure about the memory thing…). For me this is a great achievement of Vista, almost magic. I can’t tell in detail how it works (mainly because I don’t have a clue), but it works good and doesn’t cost notable performance.

So on your Vista machine you have processes that run virtualized and others that run non-virtualized. “Non-virtualized” is also called “UCA-compatible” or just “compatible”, so I’ll adopt this wording here.

You could say that Vista trusts compatible processes a little bit more than virtualized processes. That doesn’t mean virtualized processes are evil, but.. well… maybe… under certain circumstances… not 100% trustworthy. (Here is another nice article)

Note three facts here:

1. The state (virtualized/compatible) always counts for a process and is set when the process starts. When you start an .exe-file, Vista decides how to run it.
2. Because of 1) there is no sense in attaching a win32manifest to components that do not start a process, like .dlls. Thus .dlls don’t have a win32manifest.
3. The state (virtualized/compatible) of a process doesn’t change during its lifetime.

So if a .dll is used by a virtualized process, its code also runs virtualized. And if the same .dll is used by a compatible process, its code runs compatible.

Interaction

With some of that in my mind, I could explain the problems we had with our interface software arisan: arisan is a .dll you reference in your .NET-application and use it. As mentioned above arisan calls a native (so non-.NET) .dll using p/invoke-interop-techniques. Now this native .dll starts a bunch of new processes: a database-server, a database-driver, some middle-layers, etc. I am pretty sure non of this newly started processes is based on applications with a win32manifest – so they all run virtualized.

But the .NET-applications we use to test the arisan.dll started running compatible as we switched to VS2008 and the 3.5-compiler. So our test application (compatible) called arisan.dll (compatible), arisan.dll (compatible) called C-Api.dll (compatible). C-Api.dll started new processes (virtualized !!) and called functions there. And in the end a compatible (fully trusted) process asked a virtualized (less trusted) process to fill some pointers… and that’s the point where Vista interferes and says: “Dudes, are you drunk? No way!”

It was a bit of bad luck that the clash of a virtualized and a compatible process happened so deep inside a software we couldn’t debug and only threw some strange, meaningless error-messages, but on the other hand Vista could have told us more. Maybe there is some place in Vista where the clash is logged. If someone knows – please let me know, too. Let me get this straight: Vista is absolutely right here, but I’d wish more information when it happens. I don’t exactly know how, but you could do harmful things if you could easily exchange all sorts of pointers between virtualized and compatible processes…

Solution

So finally the solution was easy. Both processes have either to run virtualized or non-virtualized. Since we can’t make a third-party software run compatible, we have to make our .NET-application run virtualized: We just need to prevent .NET from inserting a win32manifest to executables.

Using the compiler from the command-line the compiler-flag /nowin32manifest does it.

In VS2008 same switch can be found on the “Application”-tab of the project properties.

The manifest-area is enabled for ‘console applications’ and ‘Windows applications’. It is disabled for class libraries.

Another way to handle this problem is to switch off UAC completely. But we strictly advice you against this solution. UAC is one of Vista’s main features and it is designed to provide shelter for your precious data in times of evil viruses and brutal intrusion attempts

Epilog- bad fix

After we understood the problem, we saw that the ‘fix’ we provided as V1.1.1 couldn’t do much. Ok – it made the demos run, because the executable compiled with the 2.0- come without a win32manifest and thus run virtualized. But users trying it with VS2008 still had the same issues.

Actually we cannot provide a technical solution with the arisan.dll, because if the running mode depends on the executables that use arisan. So we’ll write detailed information and everyone lived happily ever after.

Ok, what does this expression do or mean? First of all: never write something like this in your code. Nobody understands the meaning with a single look and that will lead to confusion and maybe to a bug in your software someday.

Now let’s take a deeper look. First you may think about the difference between x++ and ++x. If you write it as a single statement both expressions are equal, but if you use it as a part of a complex expression the first will evaluated to the value of x before increasing it; the second one will evaluate to the new value of x. So y = x++; leads to a different result for y as y = ++x;.

x+=a simply is a ‘shortcut’ for x=x+a.

Now let’s do it step by step. For example x is 5. Then first x will be increased by one to 6 but the old value will go into the formula that remains as x=x+5. Since x was increased before the result will be 11.

If you think that is all right, than please take a break and test it with your favorite compiler. If you are a C or C++ guy you will in fact receive 11 as an answer and everything is fine. But if you are a C# or java guy x will be 10. Why?

.Net as well as the Java Runtime are stack machines. The expression will be put on the stack step by step before evaluating the whole thing. At that time x is 5. x will be changed to 6 by the x++ part, but that only happens in the main memory. The old value (5) is still on the stack. After executing the whole expression the changed x will be overwritten by 10 (5+5).

And once again: NEVER write code like this!

I found this somewhere and I think it’s worth to write a short (only a very short) entry.

throw null;

If you use this expression in your C# code it will throw a NullReferenceException. That is because the throw-statement needs an object of type Exception as its single parameter. But this very object is null in my example.

I like this way to produce some exception while testing code.