What the author is missing: tail recursion in constant space is an implicit requirement for practical use of CPS transformations. In fact this is probably the single biggest reason for wanting efficient tail calls in a language.
When the continuation gets invoked by the event-handling loop, the stack is reset.
You're not compiling to C here: "the stack" may be a chain of heap-allocated activation records that are still live as far as GC is concerned. What you're describing might work but you'd want to test it to be sure.
EDIT: On reflection I agree that the setTimeout() approach should free the stack, however it's implemented.
You can always use setTimeout to call your tail, which has the neat effect that your CPS "threads" become cooperatively multitasked and can progress alongside each other. Just replace "foo()" tail calls with "setTimeout(foo, 0)".
And the downside that you're nuking your stack every time, making debugging quite a bit harder.
Twisted (a python asynchronous network programming framework) does something similar, but inverted. Twisted functions return a Deferred object without blocking. A Deferred is a promise of a value, so rather than passing the callback into the function, you register a callback with the deferred, and the function promises to call it later.
This type of thing has a few nice properties -- you can easily register more than one callback, as well as using them as a pipeline (kind-of like monads in haskell), which are awkward to implement with callbacks.
As an aside, in python you can trick generators into turning synchronous-looking code into asynchronous, which can actually make things look a lot neater if done properly -- i'll provide code if anyone's interested.
Although Diesel2 is much more interesting, with greenlets support. No more messing about with `yield`, since the blocking functions themselves force a context switch. https://github.com/jamwt/diesel
That looks really interesting, though it seems a bit dangerous -- greenlet feels like a bit of a hack.
I wish I'd found diesel before twisted; it seems a lot more elegant if you don't need any of the built-in protocols. The twisted model is good, there's just too much of it.
Suppose you wrote a compiler that translates to JavaScript in continuation-passing style. (Or suppose you wrote all your JavaScript in continuation-passing style.)
That's the implementation of call/cc.
For how to use call/cc, I recommend my other article on continuations:
I found this link before posting the question but I still can't imagine how to use javascript call/cc implementation from your article (on the other hand I think I know how to use call/cc in scheme). Maybe I am too tired today.
If you posted a simple example of JavaScript written in continuation-passing style and using this call/cc implementation probably everything would be clear.
For good descriptions of using CPS as an intermediate representation in a compiler, check out Appel's _Compiling with Continuations_, Guy L. Steele's "Rabbit" thesis, Krantz's "Orbit" thesis, and chapter 12 of the first edition of _Essentials of Programming Languages_ (they cut it from later editions, unfortunately). Appel uses ML, the rest use Scheme.
ECMA doesn't require efficient tail recursion. Caveat programmer.