Last night, after the Turing Lecture, we hosted a FlashBoF on "DSL's in Visual Studio". Stuart answered a bunch of questions and gave a much more detailed demonstration of the new DSL Toolkit than we could show during the keynote. Here's what I learned from the session:
- Models are stored in XML files. The language designer outputs an object model and will eventually also output an XSD. For example, here's a screenshot of the language designer from the DSL Toolkit we're releasing. Inside the designer, I've got a sample UIP DSL (I hacked this up on my own, this is not exactly the same one we demoed yesterday). As you can see, there's a PageCollection concept which contains Page concepts that have Name and Kind values. Page concepts also has a collection of Transfer concepts, which in turn have Label values. Generating an object model makes it easier to write tools that manipulate models. Typically, I'm anti-XML-Serialization but in this case - where we have a relative simple XSD - it works fine. I could also manipulate the model by accessing the underlying XML if I want to.
- Code generation uses templates and looks a lot like CodeSmith or old-school ASP. You interleave the static elements of the generated code with blocks of code that access the model (via the object model described above) and generate the dynamic model-specific elements of the code. So I'm guessing that people using the codegen tools like CodeSmith will feel right at home with this toolkit.
- In the current builds (which is to say later than the build that we're releasing first - the first build doesn't include any of the code generation support) we're generating a single code file from a model. Eventually we'll be able to manipulate multiple files from a single model. This is similar to how the Class Diagram works - add a new class onto the diagram and a new file gets added to the project, delete the class from the diagram and the file gets removed from the solution.
- Not all models are used to generate code. For example, in VSTS the Logical Data Center and Virtual Deployment models don't generate code. They are useful because I can use them to validate the Distributed System Model which does generate code.
- Someone asked about the implications of code coverage, profiling and test-driven development on a DSL-based process. Frankly, I don't know but it certainly got me thinking. The general consensus was that we're still in the bootstrap phase of making DSL-based development a reality and these are issues we'll have to deal with as we move forward.