CODEWITHSUNDEEP
c#metaprogrammingstatic-analysiscilopcode
Andreas Pardeike
Andreas Pardeike

Reputation: 5042

C# IL code modification - keep stack intact

This question is about static stack analysis of custom C# IL code and how to design the opcodes to satisfy the compiler.

I have code that modifies existing C# methods by appending my own code to it. To avoid that the original method returns before my code is executed, it replaces all RET opcodes with a BR endlabel and adds that label to the end of the original code. I then add more code there and finally a RET.

This all works fine in general but fails on certain methods. Here is a simple example:

public static string SomeMethod(int val)
{
    switch (val)
    {
        case 0:
            return "string1".convert();
        case 1:
            return "string2".convert();
        case 2:
            return "string3".convert();
        // ...
    }
    return "";
}

which is represented by this IL code:

.method public hidebysig static string SomeMethod(int32 val) cil managed
{
    .maxstack 1
    .locals val ([0] int32 num)
    L_0000: ldarg.0 
    L_0001: stloc.0 
    L_0002: ldloc.0 
    L_0003: switch (L_002e, L_004f, L_0044, ...)
    L_002c: br.s L_0091
    L_002e: ldstr "string1"
    L_0033: call string Foo::convert(string)
    L_0038: ret 
    L_0039: ldstr "string2"
    L_003e: call string Foo::convert(string)
    L_0043: ret 
    L_0044: ldstr "string3"
    L_0049: call string Foo::convert(string)
    L_004e: ret 
    ... 
    L_0091: ldstr ""
    L_0096: ret 
}

After my program modified it, the code looks like this:

.method public hidebysig static string SomeMethod(int32 val) cil managed
{
    .maxstack 1
    .locals val ([0] int32 num)
    L_0000: ldarg.0 
    L_0001: stloc.0 
    L_0002: ldloc.0 
    L_0003: switch (L_002e, L_004f, L_0044, ...)
    L_002c: br.s L_0091
    L_002e: ldstr "string1"
    L_0033: call string Foo::convert(string)
    L_0038: br L_009b // was ret 
    L_0039: ldstr "string2"
    L_003e: call string Foo::convert(string)
    L_0043: br L_009b // was ret 
    L_0044: ldstr "string3"
    L_0049: call string Foo::convert(string)
    L_004e: br L_009b // was ret 
    ... 
    L_0091: ldstr ""
    L_0096: br L_009b // was ret
    L_009b: my code here
    ...
    L_0200: ret
}

and I get an compile error:

Could not execute post-long-event action. Exception: System.TypeInitializationException: An exception was thrown by the type initializer for FooBar ---> System.InvalidProgramException: Invalid IL code in (wrapper dynamic-method) Foo:SomeMethod (int): IL_0000: ldnull

Is there any simple way to replace RETs in a generic way and keep the static analyzer happy?

Upvotes: 7

Views: 566

Answers (1)

Andreas Pardeike
Andreas Pardeike

Reputation: 5042

The problem turned out to be that all short jump instructions could possibly become too far away because inserting BR instead of RET increases opcode size.

I solved it by replacing all opcodes ending in "_S" with their corresponding long jump versions. For more details about this, have a look at this commit to my project: Fixed illegal short jumps

Upvotes: 3

Related Questions