I’m not a language elitist by any means. I came from a Visual Basic background and avoided C++ when I could. These days, however, I’m more of a C# guy. I feel like there is sufficient gain in using the language without much, if any development-time penalty associated with using it. While I don’t mind someone stating their reasons for using one language over the other, I will not tolerate the argument that these languages compile to the same IL and work just as well. The VB.NET and C# compilers were written by two different teams. They emitted MSIL completely different from one another by default, and with good reason on both sides. As such, there are some major differences in the IL that is output. I pointed one out the other day when talking about delegate implementation in C# vs VB.NET. Today, I’m noting another.
VB and C# Specific Language
Let’s take a look at the following scenario. I’ve created two projects in both languages and implemented the same code in both. In the form’s class, I have an array list with 100 items in it, a button control and a label control. For the button’s click event handler, I added the following code:
Private Sub Button1_Click(ByVal sender As System.Object, _
ByVal e As System.EventArgs) _
Handles Button1.Click
For i As Int32 = 1 To list.Count
Me.lblOutput.Text += list(i - 1).ToString()
Next
End Sub
This code simply loops through all of the items in the array list and outputs them to the label. Notice that I’m not caching the results of list.Count first. Instead, I’m evaluating this each time in the loop. This raises the hairs on the backs of several people’s necks every time I talk about it. I’ll explain my decision to do this later though. In the mean time, let’s look at what the C# output looks like:
for( int i = 1; i <= list.Count ; i++ ) {
this.lblOutput.Text += list[i-1].ToString();
}
Again, this whole section has the least optimal of coding decisions, but it is done in an attempt to make these two pieces of code do the same work.
The IL Comparison
These look about the same, but the problem is evident when you look at the IL. Yesterday I posted an example where the IL in VB.NET was using virtual dispatch for delegates whereas C# was using instance dispatch. Virtual dispatches are much more expensive. So what is the problem here? Let’s start by looking at the IL for the C# and VB.NET versions of this code. In C# we have a cool 34 lines of IL:
.method private hidebysig instance void button1_Click(object sender,
class [mscorlib]System.EventArgs e) cil managed
{
// Code size 64 (0x40)
.maxstack 5
.locals init ([0] int32 i)
IL_0000: ldc.i4.1
IL_0001: stloc.0
IL_0002: br.s IL_0031
IL_0004: ldarg.0
IL_0005: ldfld class [System.Windows.Forms]System.Windows.Forms.Label CSharpLoop.Form1::lblOutput
IL_000a: dup
IL_000b: callvirt
instance string [System.Windows.Forms]System.Windows.Forms.Control::get_Text()
IL_0010: ldarg.0
IL_0011: ldfld class [mscorlib]System.Collections.ArrayList CSharpLoop.Form1::list
IL_0016: ldloc.0
IL_0017: ldc.i4.1
IL_0018: sub
IL_0019: callvirt instance object [mscorlib]System.Collections.ArrayList::get_Item(int32)
IL_001e: callvirt instance string [mscorlib]System.Object::ToString()
IL_0023: call string [mscorlib]System.String::Concat(string, string)
IL_0028: callvirt instance void [System.Windows.Forms]System.Windows.Forms.Control::set_Text(string)
IL_002d: ldloc.0
IL_002e: ldc.i4.1
IL_002f: add
IL_0030: stloc.0
IL_0031: ldloc.0
IL_0032: ldarg.0
IL_0033: ldfld class [mscorlib]System.Collections.ArrayList CSharpLoop.Form1::list
IL_0038: callvirt instance int32 [mscorlib]System.Collections.ArrayList::get_Count()
IL_003d: ble.s IL_0004
IL_003f: ret
} // end of method Form1::button1_Click
But in VB.NET, take a look at the IL hit here — 44 lines of code.
.method private instance void Button1_Click(object sender,
class [mscorlib]System.EventArgs e) cil managed
{
// Code size 72 (0x48)
.maxstack 5
.locals init ([0] int32 i,
[1] class [System.Windows.Forms]System.Windows.Forms.Label _Vb_t_ref_0,
[2] int32 _Vb_t_i4_0)
IL_0000: nop
IL_0001: ldc.i4.1
IL_0002: ldarg.0
IL_0003: ldfld class [mscorlib]System.Collections.ArrayList VBLoop.Form1::list
IL_0008: callvirt instance int32 [mscorlib]System.Collections.ArrayList::get_Count()
IL_000d: stloc.2
IL_000e: stloc.0
IL_000f: br.s IL_0042
IL_0011: ldarg.0
IL_0012: callvirt instance class [System.Windows.Forms]System.Windows.Forms.Label VBLoop.Form1::get_lblOutput()
IL_0017: stloc.1
IL_0018: ldloc.1
IL_0019: ldloc.1
IL_001a: callvirt instance string [System.Windows.Forms]System.Windows.Forms.Control::get_Text()
IL_001f: ldarg.0
IL_0020: ldfld class [mscorlib]System.Collections.ArrayList VBLoop.Form1::list
IL_0025: ldloc.0
IL_0026: ldc.i4.1
IL_0027: sub.ovf
IL_0028: callvirt instance object [mscorlib]System.Collections.ArrayList::get_Item(int32)
IL_002d: callvirt instance string [mscorlib]System.Object::ToString()
IL_0032: call string [mscorlib]System.String::Concat(string,string)
IL_0037: callvirt instance void [System.Windows.Forms]System.Windows.Forms.Control::set_Text(string)
IL_003c: nop
IL_003d: nop
IL_003e: ldloc.0
IL_003f: ldc.i4.1
IL_0040: add.ovf
IL_0041: stloc.0
IL_0042: ldloc.0
IL_0043: ldloc.2
IL_0044: ble.s IL_0011<
IL_0046: nop
IL_0047: ret
} // end of method Form1::Button1_Click
There are 10 additional IL lines for the same exact code — four of which appear to be unnecessary “nop” fields. The MSDN help for nop states: “Fills space if opcodes are patched. No meaningful operation is performed although a processing cycle can be consumed.” So despite the fact that these instructions do nothing, they can be wasting important processing cycles, which, if executed in a loop such as ours can be quite expensive over the long run.
Let’s break this down a bit further to see exactly what’s happening in these functions. Let’s evaluate the C# side first so I can pick on VB.NET’s IL later and show you where it’s going, well, wrong 🙂
C# IL Code Analysis
First off, the C# header is pretty standard.
.method private hidebysig instance void button1_Click(object sender,
class [mscorlib]System.EventArgs e) cil managed
{
// Code size 64 (0x40)
.maxstack 5
Next we see our locally scoped declarations. In this instance, we only have one variable that is initialized as an int32. This will be the variable that we use in our loop counter.
.locals init ([0] int32 i)
We then push an int32 value of 1 onto the evaluation stack and immediately pop it back off into the local variable we declared above.
IL_0000: ldc.i4.1 IL_0001: stloc.0
The next step is fairly interesting for anyone that hasn’t seen IL before. We are transferring control to the instruction at the IL_0031 label. We’ll cover this in a minute.
IL_0002: br.s IL_0031
We are now going to load the first argument onto the evaluation stack (stack size +1). This happens to be the “sender” parameter of the click event. We use the ldfld call to push the label on the stack. The IL is just preparing the text value to be concatenated in the loop.
IL_0004: ldarg.0 IL_0005: ldfld class [System.Windows.Forms]System.Windows.Forms.Label CSharpLoop.Form1::lblOutput
Having fun so far? Good, let’s keep on trucking. Next, we are doing something fairly interesting. We are going to duplicate topmost item on the evaluation stack and push it onto the stack as well. The way this happens is fairly complicated for such a small field name. First, the value is pushed onto the stack, popped off the stack and duplicated, then pushed back onto the stack. Once we’ve performed three operations to get that done, we perform one more to push the duplicated value on the stack. Remember out += operator in the code? Sound like something a += operator might want to do? I hope so :). Next, we are going to use callvirt to push the return value of the label’s Text property onto the evaluation stack.
IL_000a: dup IL_000b: callvirt instance string [System.Windows.Forms]System.Windows.Forms.Control::get_Text()
We’ve seen the next series before so I’ll spare you the detailed explaination. We are pushing our array list onto the stack. We need this so we can get our count later.
IL_0010: ldarg.0 IL_0011: ldfld class [mscorlib]System.Collections.ArrayList CSharpLoop.Form1::list
Remember way back at the start of this method, we initialized a local variable? Well now we are going to load it. We then pushes an int32 value of 1 onto the evaluation stack again.
IL_0016: ldloc.0 IL_0017: ldc.i4.1
We do our subtraction that is done in the array indexer (“i-1”).
IL_0018: sub
Next we use callvirt again to get the item at the index provided by the result of the sub field’s return value. Without delay, we call that object’s ToString method and store the return value on the stack.
IL_0019: callvirt instance object [mscorlib]System.Collections.ArrayList::get_Item(int32) IL_001e: callvirt instance string [mscorlib]System.Object::ToString()
Of course, our call is concatenating strings so we call the String.Concat method using our Control’s text value again. Notice when we called concat we are using “call” instead of callvirt? This is because we know the object is a native type, String. We used callvirt against our controls because these objects can be recast at run-time. Therefore the callvirt method is used to call the method against the runtime type of the object, not the compile time type.
IL_0023: call string [mscorlib]System.String::Concat(string,string) IL_0028: callvirt instance void [System.Windows.Forms]System.Windows.Forms.Control::set_Text(string)
Here we go again with this sequence! This time we are loading the values from our first parameter so we can increment the value with the “add” field. This adds the value 1 to the value stored at location 0. Make sense? Good, because we then have to call stloc again to store the value back into location 0.
IL_002d: ldloc.0 IL_002e: ldc.i4.1 IL_002f: add IL_0030: stloc.0
Now we have to load the value stored in our first initialized value again so we can do our “for(;;)” comparison. We compare this value against the value returned from the ArrayList’s Count method.
IL_0031: ldloc.0 IL_0032: ldarg.0 IL_0033: ldfld class [mscorlib]System.Collections.ArrayList CSharpLoop.Form1::list IL_0038: callvirt instance int32 [mscorlib]System.Collections.ArrayList::get_Count()
Remember back near the beginning of all of this, we transferred control to an instruction at IL_0031 ? Well now we are transferring control back to IL_0004. If this sounds like loop control instructions to you, you’d be right. Lastly, this method ends with a return call. Since this method is void, we don’t return any values.
IL_003d: ble.s IL_0004 IL_003f: ret } // end of method Form1::button1_Click
Was this code really necessary for a simple loop!? Sadly, yes. But before you get upset, lets remember that we have 10 extra lines in our VB.NET compiled IL! I wont rehash all of the VB.NET code like we did with C#, but I will concentrate on the differences because that’s what causes our performance hit in VB.NET.
VB.NET IL Code Analysis
Take a look at the first difference — the initialized variables.Notice that we have 3 variables declared here instead of 1. The first is the variable that we use in the loop, and the second is a reference to the label that we will use in the loop. The third is used to cache the value of the Count() method later on. Rather than evaluate this within the loop, we will cache this value in a local variable and use this value in our loop comparison.
.locals init ([0] int32 i,
[1] class [System.Windows.Forms]System.Windows.Forms.Label _Vb_t_ref_0,
[2] int32 _Vb_t_i4_0)
The next thing I’ll touch on is those nasty nop calls. These are annoying but are there for debugging support. These allow you to place break points in code that have no execution. These are just placeholders, but as said, they do waste instructions. Try compiling these in release mode instead of debug mode to get rid of these. You should be doing this anyway when you are testing the performance of your code.
The IL that follows is representative of what we’ve already said. The label will be loaded as a local variable along with a cached value of the variable. These values are obtained and stored in location 0 and 2 respectively. (stloc.0 and stloc.2 fields)
L_0001: ldc.i4.1 L_0002: ldarg.0 L_0003: ldfld [mscorlib]System.Collections.ArrayList VBLoop.Form1::list L_0008: callvirt instance int32 [mscorlib]System.Collections.ArrayList::get_Count() L_000d: stloc.2 L_000e: stloc.0
We do the same loop control we had before with the br.s call and then move in to load the first argument (our label) and put it on the stack. We load the value stored in the local variable stored in location 1 twice (ldloc.1). Notice this is different than the C# version of “dup”.
L_0011: ldarg.0 L_0012: callvirt instance [System.Windows.Forms]System.Windows.Forms.Label VBLoop.Form1::get_lblOutput() L_0017: stloc.1 L_0018: ldloc.1 L_0019: ldloc.1
Moving forward we get our text value of the output label and start our loop through the array list.
L_001a: callvirt instance string [System.Windows.Forms]System.Windows.Forms.Control::get_Text() L_001f: ldarg.0 L_0020: ldfld [mscorlib]System.Collections.ArrayList VBLoop.Form1::list
We load our first local variable and do our subtraction again. But notice our subtraction call is different.
L_0025: ldloc.0 L_0026: ldc.i4.1 L_0027: sub.ovf
This version of sub.ovf is used to do overflow checking on the subtraction operation. This is turned on by default in VB.NET compilation, but turned off in C# compilation.
We take the same steps to load the item in our array list, get the result of the ToString call, concatenate the value, and then set the text value of our label that we did in the C# version.
L_0028: callvirt instance object [mscorlib]System.Collections.ArrayList::get_Item(int32) L_002d: callvirt instance string object::ToString() L_0032: call string string::Concat(string, string) L_0037: callvirt instance void [System.Windows.Forms]System.Windows.Forms.Control::set_Text(string)
As we draw near the end of our IL, again go through the steps to increment our loop counter. After we put some more of those annoying little nop calls. Just as the subtraction call did overflow checking, the add method does overflow checking, hence the change from add to add.ovf.
L_003c: nop L_003d: nop L_003e: ldloc.0 L_003f: ldc.i4.1 L_0040: add.ovf L_0041: stloc.0
We end our IL nicely by reloading our variables in position 0 and 2 before returning to the front of our loop.
L_0042: ldloc.0 L_0043: ldloc.2 L_0044: ble.s L_0011 L_0046: nop L_0047: ret }
Summary
This has really just been an informational look at how the compilers differ by default. Its entirely possible to change a few parameters in the compilers to get the output to look a bit more alike. Its important to know these differences, though, so that you can make educated decisions on how you should optimize your code based on what language you are working with. Happy hacking.
