Designing a Domain-Specific Language

Designing an embedded language requires two steps: first, design the language that’ll allow you to express the things you want to express, and second, implement a processor, or processors, that accepts a “program” in that language and either performs the actions indicated by the program or translates the program into Common Lisp code that’ll perform equivalent behaviors.

So, step one is to design the HTML-generating language. The key to designing a good domain-specific language is to strike the right balance between expressiveness and concision. For instance, a highly expressive but not very concise “language” for generating HTML is the language of literal HTML strings. The legal “forms” of this language are strings containing literal HTML. Language processors for this “language” could process such forms by simply emitting them as-is.

(defvar *html-output* *standard-output*)
(defun emit-html (html)
  "An interpreter for the literal HTML language."
  (write-sequence html *html-output*))
(defmacro html (html)
  "A compiler for the literal HTML language."
  `(write-sequence ,html *html-output*))

This “language” is highly expressive since it can express any HTML you could possibly want to generate.1 On the other hand, this language doesn’t win a lot of points for its concision because it gives you zero compression—its input is its output.

To design a language that gives you some useful compression without sacrificing too much expressiveness, you need to identify the details of the output that are either redundant or uninteresting. You can then make those aspects of the output implicit in the semantics of the language.

For instance, because of the structure of HTML, every opening tag is paired with a matching closing tag.2 When you write HTML by hand, you have to write those closing tags, but you can improve the concision of your HTML-generating language by making the closing tags implicit.

Another way you can gain concision at a slight cost in expressiveness is to make the language processors responsible for adding appropriate whitespace between elements—blank lines and indentation. When you’re generating HTML programmatically, you typically don’t care much about which elements have line breaks before or after them or about whether different elements are indented relative to their parent elements. Letting the language processor insert whitespace according to some rule means you don’t have to worry about it. As it turns out, FOO actually supports two modes—one that uses the minimum amount of whitespace, which allows it to generate extremely efficient code and compact HTML, and another that generates nicely formatted HTML with different elements indented and separated from other elements according to their role.

Another detail that’s best moved into the language processor is the escaping of certain characters that have a special meaning in HTML such as <, >, and &. Obviously, if you generate HTML by just printing strings to a stream, then it’s up to you to replace any occurrences of those characters in the string with the appropriate escape sequences, <, > and &. But if the language processor can know which strings are to be emitted as element data, then it can take care of automatically escaping those characters for you.