[racket] iterative lazy space-safe or not?

From: Keiko Nakata (keiko at kurims.kyoto-u.ac.jp)
Date: Thu Nov 18 09:24:31 EST 2010

> Well, think about `delay' as it's used in most Scheme implementations:
> you use it in specific contexts where usually you know in advance that
> you're not worried about exceptions -- if an exception always
> indicates a bug, then it doesn't matter that forcing some promise
> throws an exception, and a later force doesn't -- either way it's a
> bug that you need to resolve, and when the bug is gone you're back
> into "don't care about exceptions" mode.
> 
> But if you want to implement a lazy language, then the perspective
> changes, and you certainly want that consistent behavior.  Same as
> racket itself -- if exceptions *always* indicate a bug in your code
> then there's no real need for making up the machinery to capture them
> etc, you can get away with reporting the problem and then segfault
> your way out of the process.

I am interested in lazy evaluation in the context of effectful
call-by-value languages and I do care consistent behavior. Also, I am
not convinced if I'll be willing to crash the system when
initialization fails or throw uncaught exceptions...

This said, if I commit to that initialization is performed only once,
raising exceptions for black holes and memorizing exceptions seem
only natural. Well, I could instead silently diverge!

> But a bit more than that, I made `lazy' promises usable as the end
> points also, which means that you can use them too -- provided that
> you don't need multiple values.  In fact, I think that in general
> using `lazy' is better since it composes -- it's in perfect analogy to
> tail calls: it's true that you don't always need them, but by getting
> them wherever possible "good things happen", including situations
> where you didn't realize that you need it.  So if you're using `delay'
> you're making sure that promises don't compose, whereas a consistent
> use of `lazy' will give you that when needed.

But 'lazy' could incur tag-check twice when 'delay' does just once?
Also 'lazy' may involve two indirections when 'delay' does one?
I.e., I thought there is trade-off.

Keiko



Posted on the users mailing list.