...
The code generation templates and the runtime library itself are essentially feature complete with respect to the reference Java version. The C# (and Python) targets have a definite advantage due to the existence of a native StringTemplate implementation. There are still currently zero automated tests written and it isnthere hasn't yet clear been much time to investigate how such tests can/should be written (we suspect this applies to all the other targets too). In any case, some basic sanity check are done with samples grammars in the examples-v3 archive.
UPDATE: Just learned that some Some of the other targets apparently have some automated unit tests already. We'll have a peek and see if this can be applied to C# Hopefully this can suggest an approach that is applicable to the C# target too.
Architecture
The As with all other targets, the C# code generation and runtime are based modelled on the Java version and would remain so until well after the final release. ANTLR 3 development (including AntlrWorks debugging support and some of the internal interfaces) is still so fluid that investing significant effort in developing radically different code output would be a little premature at this stage. That would be . This means the C# target supports features such as grammar development/prototyping and remote debugging with the AntlrWorks GUI which is very important for ANTLR users. Developing radically different code output is left as an exercise to the reader 
...
Microsoft Visual Studio 2003 and 2005
Nant v0.85
Performance
Should be The V3 target generates code that is easily faster than V2 C# code generationthat generated by the V2 target (especially the lexers). We probably won't be able to outrun match the bare-metal performance of the code generated by Jim Idle's C target output or Ric Klaren's C++ target but, we expect to be very competitive with the other targets.
...