MSIL De-Compiler Working Mechanism

Introduction

Welcome to my journey of writing a .NET assembly de-compiler. First of all I’ll try to develop a theory to decompile MSIL. I just do whatever a MSIL instruction ask me to do. But I do it keeping in mind that I am decompiling MSIL. So when it asks to push me a value of a variable I push the name of that variable on stack.
Simple Case Example

To understand the code it is required that you know or have a reference of what each instruction of MSIL actually does. Here is sample program to test if our concept works:
namespace DisasmIL
{
class Math
{
public int add(int x, int y)
{
return x + y;
}
}
class Program
{
static void Main(string[] args)
{
Math m;
int a, b;
m = new Math();
a = 20;
b = 50;
int p = m.add(a, b);
}
}
}

We only check one method. The Main method.When the Main method is
compiled it takes following form.
.method private hidebysig static void Main(string[] args) cil managed
{
.entrypoint
// Code size 23 (0x17)

.maxstack 3
.locals init (
[0] class DisasmIL.Math m,
[1] int32 a,
[2] int32 b,
[3] int32 p
)
IL_0000: nop
IL_0001: newobj instance void DisasmIL.Math::.ctor()
IL_0006: stloc.0
IL_0007: ldc.i4.s 20
IL_0009: stloc.1
IL_000a: ldc.i4.s 50
IL_000c: stloc.2
IL_000d: ldloc.0
IL_000e: ldloc.1
IL_000f: ldloc.2
IL_0010: callvirt instance int32 DisasmIL.Math::’add'(int32,int32)
IL_0015: stloc.3
IL_0016: ret
} // end of method Program::Main
We parse line by line:
.method private hidebysig static void Main(string[] args) cil managed

It is method declaration with default starting curly brace. Output code:
static void Main(string[] args)
{

Stack: [empty]
.entrypoint
// Code size 23 (0x17)

.maxstack 3
.locals init (
[0] class DisasmIL.Math m,
[1] int32 a,
[2] int32 b,
[3] int32 p
)

Need not to be explained. They are self explanatory. Declare variables.Output code:
DisasmIL.Math m;
int a;
int b;
int p;
Stack: [empty]
IL_0000: nop:

Does nothing (nop).
Output code: [none]
Stack: [empty]
IL_0001: newobj instance void DisasmIL.Math::.ctor()

Create new instance of DisasmIL.Math using default constructor
so we push “new DisasmIL.Math()” on our stack.
Output code: nothing.
Stack: new DisasmIL.Math()
IL_0006: stloc.0

So we pop top of stack and assign it to local variable 0.Output code:
m = new DisasmIL.Math();

Stack: [empty]
IL_0007: ldc.i4.s 20

What we do is push constant 20 on stack.
Output code: [none]
Stack: 20
IL_0009: stloc.1

So we pop top value and assign it to local variable 1. Output code:
a = 20;

Stack: [empty]
IL_000a: ldc.i4.s 50

We push constant 50 on stack.
Output code: [none]
Stack: 50
IL_000c: stloc.2

We pop top value and assign it to local variable 2.
Output code:
b = 50;

Stack: [empty]
IL_000d: ldloc.0
IL_000e: ldloc.1
IL_000f: ldloc.2

Push local variable 0, 1 and 2 on stack.
Output code: none.
Stack: m, a, b
IL_0010: callvirt instance int32 DisasmIL.Math::’add'(int32,int32)
IL_0015: stloc.3

We call add method with values top-1, top of stack for instance of top-2. For any method call if it returns value it is returned on stack. So check next instruction. If it is a stloc then we assign the return value. We assign return value to local variable 3.

Output code:
p=m.add(a,b);

Stack: [empty]
IL_0016: ret

Return void. So no code except closing curly brace.Output code:
}

Stack: [empty]

Now if you add the output codes together you’ll find the original C# code is generated. This works for simple cases. Need to test if it works for complex situations. Lets do some more interesting thing now.
The if Structure

Lets se at the most basic and useful if-then structure. The compiler generates a conditional branch. We create a block of instructions for the structure. The block is separated by the branch instruction (like brtrue) and the branch label (like IL_0019 – where the code jumps). And our condition is on the stack. If we find a true condition branch we negate it and put it as a if statement condition. The block is initially a block of MSIL that we will convert to C# code later (may need recursion here??). Please note that the labels are not stored in MSIL. It is just the byte offset of the MSIL in a method.

We take a simple method to test our theory with if structure.
public int IfStructure(int a, int b)
{
if (a < b)
{
System.Console.Write(“Condition is true”);
}
return b;
}
Here is the MSIL code generated by Visual Studio 2005 compiler.
.method public hidebysig instance int32 IfStructure(int32 a, int32 b) cil managed
{
// Code size 31 (0x1f)

.maxstack 2
.locals init ([0] int32 CS$1$0000,
[1] bool CS$4$0001)
IL_0000: nop
IL_0001: ldarg.1
IL_0002: ldarg.2
IL_0003: clt
IL_0005: ldc.i4.0
IL_0006: ceq
IL_0008: stloc.1
IL_0009: ldloc.1
IL_000a: brtrue.s IL_0019
IL_000c: nop
IL_000d: ldstr “Condition is true”
IL_0012: call void [mscorlib]System.Console::Write(string)
IL_0017: nop
IL_0018: nop
IL_0019: ldarg.2
IL_001a: stloc.0
IL_001b: br.s IL_001d
IL_001d: ldloc.0
IL_001e: ret
} // end of method ControlStructures::IfStructure
We now start parsing now-
.method public hidebysig instance int32 IfStructure(int32 a, int32 b) cil managed
{
// Code size 31 (0x1f)

.maxstack 2
.locals init ([0] int32 CS$1$0000,
[1] bool CS$4$0001)

These lines generate output that we do without any processing of MSIL. There is method definition and local variables. We change the local variable names to C# current names without conflict. For simplicity here we just replace ‘$’ with ‘_’. One thing to evaluate the MSIL instructions we must keep a map of local variables with variable number. In this example CS$1$0000 is local variable 0 of type int. For clarity we do not show the map here. A simple STL map should work.

Output:
public int IfStructure(int a, int b)
{
int32 CS_1_0000;
bool CS_4_0001;

Stack: [Empty]
IL_0000: nop
IL_0001: ldarg.1
IL_0002: ldarg.2

So we push method argument 1 and 2 to stack:

Output: [None]

Stack: a,b
IL_0003: clt

This instructs us if stack top-1 is less than stack top. The two elements are popped from stack and result goes to stack. No output of course.

Output: [None]

Stack:
a<b
IL_0005: ldc.i4.0

Load (means push) constant integer of value 0 on stack.

Output: [None]
Stack: a<b,0
IL_0006: ceq

Check if stack top-1 equals stack top. Result goes to stack.
Output: [None]
Stack: a < b == 0
IL_0008: stloc.1
IL_0009: ldloc.1

What else? Store stack top in local variable 1 and load that on tack again. We decided previously when we store some value in a local variable we use assignment to that variable and output that code. For clarity I added parentheses:

Output:
CS_4_0001 = (a < b == 0)

Stack:CS_4_0001
IL_000a: brtrue.s IL_0019

We have got a conditional branch. We create a block starting from here to IL_0019. And put them in curly braces. And our condition is on the stack. We find a true condition branch so we negate it and put it as if structure as I said at the beginning.

Output:
if(!CS_4_0001)
{
IL_000c: nop
IL_000d: ldstr “Condition is true”
IL_0012: call void [mscorlib]System.Console::Write(string)
IL_0017: nop
IL_0018: nop
}

Stack: [Empty]
IL_0019: ldarg.2
IL_001a: stloc.0
IL_001b: br.s IL_001d
IL_001d: ldloc.0
IL_001e: ret

We do not parse them here. They are very simple to understand and we can parse them using the method we seen in simple example case.

OK, we now can work on little more complex codes than easiest. This will also produce codes that was generated by “for structure” but in a funny way. If we add a little more intelligence to produce “goto” output code for special branching that we cannot handle with if, we get following result.

The code like this:
for(int i=0;i<10;i++)
{

}

will be converted to:
int i;
i=0;
label_1:

if(i<10)
{

i++;
goto label_1;
}

OK, but no problem. We’ll look at loops later. Now you may find that our theory generates a funny code block like:
CS$4$0001=((a<b)==0);
if(!CS$4$0001)
{

}

Here the optimization comes to scene. But we skip them now.

Loops

I wish, as a decompiler writer, there would be no loops. Programmers use thousands of goto statements with if statements. But as it is not the case I must understand how to parse MSIL instructions that were generated from the loops.

Of the three types of most common loops (for, while, do-while) the while loop is the basic one. The block that is generated from any type of these loops usually has a conditional jump (usually a brtrue.s ) as the last instruction of the block. The difference from if structure is- the instruction jumps to an offset less than current instruction offset. The for and while loop has a unconditional branch (br.s ) to an offset that is between start and end of the block. The jump target usually at the beginning of the condition checking instructions. The do-while loop lacks this branch for the reason – it does not test the condition before it is at the end of the block.

So, we get instruction block like following MSIL block:
IL_0010: br.s IL_005a ;do-while loop does not have this line

IL_0012: nop

[any type and number of instructions]

IL_0059: nop

[condition check instruction- results boolean value on stack]

IL_0060: brtrue.s IL_0012

While looking at if structure we have seen how to create a boolean condition for if structure. Things are similar here for the loops. Follow the instructions – get the top stack element when conditional jump found – reverse it (add just an !) for brtrue.s jump and put it as the loop statements condition statement. Please note that the conditional jump targets the instruction just at or after the starting instruction of the block.

Here we find we can not have a single passing decompiler. We must identify the code blocks in an iteration before final iteration. Till now we can identify blocks of if, for, while, do-while structures by using conditional jump instructions and their destination. If has destination offset after the current offset and others have destination before the current offset. The for and while can not be distinguished very clearly but the do-while does not have a jump at the beginning. And of course there can be nested blocks that are generates from nested loops.

There is some complex variation of the loops – like infinite loops, forcach loop etc. They are not much different. But for now let me keep that unfinished and wait for your response to fix errors and continue further then. Comment on it, vote for it and show me my mistakes.