Transparent Compilation

12 February 2013

Increasingly web developers are using languages like CoffeeScript and SCSS that compile to other textual source languages that execute in the browser. Such source-to-source compilers (also called transpilers 1) are not new, Cfront was widely used in the early days of C++ to generate target C code. But for me there is a difference that picks out CoffeeScript and SCSS as transparent compilers

1: Hunting around usage, it seem to me that the the term “transpiler” is used as a synonym for source-to-source compiler. So transpilers may or may not be transparent. I've also seen the term “source-to-source translation” used equivalently to “source-to-source compilation”.

With most compilers, you don't care very much about what is generated downstream. As long as it follows the semantics of the source language it's effectively a big lump of bits. But if you're going to generate JavaScript for the browser, this ignorance is hard to live with. Debugging environments are getting pretty nifty these days, but they are all in terms of the HTML/CSS/JavaScript triad. So it's important that you understand how your input language translates to its executable target.

This constraint has a big effect upon the source language. You need to ensure that the output corresponds very clearly to the source. When I write this CoffeeScript

$(window).on 'touchTap', (event) ->
  window.touchPanel.tap(event)

I can easily recognize the resulting JavaScript in my browser debugger

$(window).on('touchTap', function(event) {
  return window.touchPanel.tap(event);
});

This remains true even for more complicated bits of CoffeeScript's transformation, such as turning

runSetupBuild: (slide, positionClass) ->
  switch positionClass
    when 'current', 'next'
      @buildsFor(slide)?.setupBuild?.forwards()
      # ...

into

Infodeck.prototype.runSetupBuild = function(slide, positionClass) {
  var _ref, _ref1, _ref2, _ref3;
  switch (positionClass) {
    case 'current':
    case 'next':
      return (_ref = this.buildsFor(slide)) != null ? 
           (_ref1 = _ref.setupBuild) != null ? _ref1.forwards() : void 0 
           : void 0;
    /* ... */

There's a lot going on in that transformation - but still the correspondence is pretty clear. If I need to debug in that code, I can easily see how it matches up to the source CoffeeScript. This is the essence of what makes the compilation process transparent - the intention that you will work in the output language 2.

2: Even with opaque compilation there are cases when people study the outputs. Occasionally there are odd behaviors or bugs that do require you to dig into the compiler outputs. Some programmers like to understand what the compiler is doing, although that's got less common as compilers and virtual machines get more sophisticated. But such activity is an exception.

In contrast, there are source-to-source compilers that don't expect you to work in the output language, or see visibility of the output language as an unfortunate temporary mechanism. These can still be useful, and you see them in the javascript world with languages like Dart, GWT, and ClojureScript. It's this difference in intention which is what separates the transparent style of transpilation from the more common approach. 3

3: It's interesting to see whether the development of sourcemaps will shift languages like coffeescript away from transparency.

The fact that you have to work at keeping the compilation transparent puts limits on what you can do in the source language. You don't have the degree of freedom in your language constructs that you get with a more unconstrained form of compilation. You have to follow the basic semantics of the target language, and keep to much the same program structure. These constraints haven't been widely talked about as a feature of language design.

CoffeeScript illustrates that despite these constraints you can get a considerable difference in syntax between source and target languages. CoffeeScript feels much more like Python in syntax than the C-like JavaScript. Such syntactic variation isn't always the case, indeed there is an important subset of transparent compilation languages that strive to be a superset of the target language. SCSS and TypeScript fit into this category - any CSS expression is valid in SCSS. Using a superset language makes the correspondence even clearer and I feel works well for CSS where CSS's syntax works well but the language misses some handy features.

Some say there's little point in using transparent compilation - if you have to understand the target code for debugging, what's the value in using a different source? For me, the value lies in a couple of directions. First off it's a way of getting useful language features that are missing in the target language. SCSS gives me handy capabilities such as variables (so I can say $light-purple instead of #f8c8fe and change it in only one place should I want to tweak it).

More drastic syntax changes, such as CoffeeScript, require stronger justification. One of my colleagues put it very well after finishing a project. He's an experienced JavaScript programmer and the project wrote well-disciplined JavaScript from the beginning. As a result he was very happy with the quality of the JavaScript codebase. However he still concluded that they would have been better off working in CoffeeScript, because it's easier to understand what is going on when you're reading the CoffeeScript, even when you're debugging in the generated JavaScript code. The transformation may not look like such a big deal for small fragments, such as those I show above. But it makes a big difference once you're up to hundreds of lines of code, let alone beyond.4

4: I've only done a few hundred lines of coffeescript while working on my infodecks, but I agree with him and will continue to use coffeescript for any non-trivial amounts of javascript.

Notes

1: Hunting around usage, it seem to me that the the term “transpiler” is used as a synonym for source-to-source compiler. So transpilers may or may not be transparent. I've also seen the term “source-to-source translation” used equivalently to “source-to-source compilation”.

2: Even with opaque compilation there are cases when people study the outputs. Occasionally there are odd behaviors or bugs that do require you to dig into the compiler outputs. Some programmers like to understand what the compiler is doing, although that's got less common as compilers and virtual machines get more sophisticated. But such activity is an exception.

3: It's interesting to see whether the development of sourcemaps will shift languages like coffeescript away from transparency.

4: I've only done a few hundred lines of coffeescript while working on my infodecks, but I agree with him and will continue to use coffeescript for any non-trivial amounts of javascript.