Skip to main content

A DSL in 36 lines of code

· 4 min read
Tijs van der Storm

One of the goals of Rascal is to allow the definition of Domain-Specific Languages. In this small post we give a flavor of how you can use Rascal to define the syntax of a DSL, a simple semantic check and how to compile the DSL to Java.

The following example shows how to define a simple DSL for state machines. It includes a parser, a check for unreachable states and a compiler to Java code.

The grammar of the DSL is defined using Rascal's grammar formalism which is fully integrated in the language. Shown below is the syntax definition of a simple state machine language, inspired by Martin Fowler's example language for gothic security.

module Syntax

extend lang::std::Layout;
extend lang::std::Id;

start syntax Machine = machine: State+ states;
syntax State = @Foldable state: "state" Id name Trans* out;
syntax Trans = trans: Id event ":" Id to;

A state machine consists of a number of named state declarations, where each state contains transitions to other states (identified by name) when a certain event happens. The grammar reuses identifier syntax and whitespace convention from the standard library. Each non-terminal defines a type. Parse trees are typed values like any other value in Rascal. As a result, you can write functions that process such trees. An example would be a semantic check on state machines, such as finding all unreachable states:

module Analyze

import Syntax;

set[Id] unreachable(Machine m) {
r = { <q1,q2> | (State)`state <Id q1> <Trans* ts>` <- m.states,
(Trans)`<Id _>: <Id q2>` <- ts }+;
qs = [ | q &- m.states ];
return { q | q <- qs, q notin r[qs[0]] };

To check for unreachable states, we first create a binary relation between states using a comprehension. This comprehension uses concrete syntax matching to find a state's transitions. The pattern between backticks is written in the object language, which in this case is the statemachine language defined in the grammar above (Note the embedded syntax highlighting!). The variables q1 and ts in between < and > are bound for each state that is found in the machine m. A similar pattern is used to find the target state q2 is found in each transition in ts. The post-fix + then computes the transitive closure of the relation.

The relation r is based on the transitions in a state machine. This means that it does not include declared (final) states which have no outgoing transitions. We collect the names of all defined states in qs , again using a comprehension.

The initial state is (conventionally) defined to be the state that is declared first. An unreachable state is then defined as a state that is not in the right image of the initial state in the transitive closure of the transition relation. This is exactly what is described by the last comprehension! The notation r[x], where r is a relation is short hand for { y | <x, y> <- r }.

There are various ways of compiling a DSL to target code in Rascal. The simplest is using string templates and generate code in a general purpose language. The following snippet shows the generation of a Java while loop to execute a state machine.

module Compile

import Syntax;

str compile(Machine m) =
"while (true) {
' event =;
' switch (current) {
' <for (q <- m.states) {>
' case \"<>\":
' <for (t <- q.out) {>
' if (event.equals(\"<t.event>\"))
' current = \"<>\";
' <}>
' break;
' <}>
' }

String templates allow arbitrary Rascal values and control-flow constructs to be interpolated in string literals. Note how this code does not use concrete matching, but instead uses the labels defined in the grammar (i.e., states, out, event, and to).

And that's it! A complete DSL in 36 lines of code. Of course, the parser and the unreachable and compile functions can be connected to the IDE. This provides custom syntax highlighting, error-marking and automatic building in state machine editors.