dikini.net

I've been looking more into haskell, as part of some personal projects. There I need some constant array like structures, that is ones which have integer indexing and there are some long winded reasons I'm not using a Map. As a first step I've decided look at haskell arrays, by writing some search functions. It is mostly a syntax exercise.

I'm on a learning haskell journey. But reading someone else's code and thoughts while educational is not that enlightening. Anyway, to cut the long story short:

• I (still) want to see what this haskell thing is all about
• I need a moderately complex problem to get a feeling of the haskell type system works, face the dreaded IO, while not wasting too much time
• I need a decent actor style abstraction for some social/population based methods experiments I'm into

So this is how ended with Why not try my hand at erlang style concurrency. The best thing about it is that it aligns nicely with an individualistic view of the world. I can write sequential code to do something, which from time to time communicates with the rest of them, while the rest of the time doesn't really care.

Erlang implements lightweight processes with asynchronous messaging. The cool twist it puts on it is having reception guards - if a message is not matched by the guard conditions it stays in the mailbox (channel), and the rest of the messages are scanned. The first matching message is removed from the mailbox, the remaining messages are left in the mailbox in arrival order.

control resumed inspired me to take a small exercise.

"Code a minimal approximation of the python yield operator based generators"
Here follow the results I managed to get with help from Matthias Felleisen, via the plt mailing list. You can see the non-edited exchange via the mailing list archives and control and macors. All control&prompt are forms specific to plt scheme's control.ss module. The first attempt looks something like this:

(require (lib "control.ss"))

(define (make-step)
  (define (control-state)
    (yield 1)
    (yield 2)
    (yield 3)
    'finished)
  (define (yield value) (control resume-here (set! control-state resume-here) value))
  (define (generator) (prompt (control-state)))
  generator)

(define step (make-step))

(step) (step) 

It does the job. Kind of. For me the interesting bit is how the interplay of prompt and control achieve maintaining the current state of execution before the return and allowing to continue from that interruption. It can be done with call/cc but it looks a helluva a lot more complicated and is far more unreadable.

There is a problem with the code above though. As Matthias notes in his email:

LESSON: never 'test' by running things at the top-level prompt. It IS a prompt. -- Matthias

So comes version 2:

(require (lib "control.ss"))

(define (make-step)
  (define (control-state)
    (yield 1)
    (yield 2)
    (yield 3)
    'finished)
  (define (yield value) 
    (control resume-here 
             (set! control-state 
                   (lambda () (prompt (resume-here)))) 
             value))
  (define (generator) (prompt (control-state)))
  generator)

(define step (make-step))

(equal? '(1 2 3) (list (step) (step) (step)))

control-state is a variable holding a reference to a function with no arguments, doing what the python function would do. yield is a variable holding a reference to a function which does the interrupt magic. generator is a variable holding a reference to a (sugary) function which does the initial step.

The logic of the above algorithm is roughly as follows - mark thy state, do what you need to do, when you encounter control stop, remember your state by wrapping it in a function (resume-here), remember where do you what to continue from ( the set! control-state ...). Notice that the value of control-state is replaced. This is possible since the resume-here continuation is a function and a first class value.

So what next. Boilerplate code is tedious to write. And there is a lot of boilerplate code there. And every function, which needs to follow this recipe, will need to duplicate it.

So scheme macros to the rescue. It is not a trivial problem. A 'generator' function can look something like this:

Here we go again. In this writeup I'll revisit the ideas of closures, state and closures, partial evaluation and generics.

Basics of 'functions as first class language objects'

Yep. PHP can assign functions to variables. In this sense functions are first class objects. Object and class methods are similar, you can, kind of, assign them to variables and later execute them. What does this change? Well, it improves our ability to abstract common patterns and idioms. There are examples of this in the above writeups as well as in the aspects one and quite a few of the rest of php ones on this site. The basic concept is php callback for example: call_user_func('a_callback_function'); call_user_func(array('AClass', 'aCallback')); call_user_func(array($obj, 'aCallback')); $callback = 'a_callback_function'; $callback(); $callback = array($obj, 'aCallback'); $callback(); What can we use this for? The most obvious and straight forward this is a lightweight state pattern implementation - $a_thing() Will behave differently depending on the value of the $a_thing variable, it's state. Another way to look at php callbacks is as proxies for 'real' functions, class or object methods. Oh, let's not forget is_callable, it's a very useful little number.

Objects and functions

The facility of PHP objects to encapsulate a state and us being able to access this state later can be used to turn them into functions. Let's consider: class aClass{ var $state; function get() { return $state; } function next() { $state=random; } function callbacks() { return array( array($this,'get'), array($this, 'next'));} } $obj = new aClass(); list( $random, $next_random ) = $obj->callbacks(); echo $random(); echo $next_random();

Here comes a sample test suite (it is too long, so only the beginning features in here) for the any/types, I described earlier. I like the clenliness of the schemeunit library. It features a rich set of assertions, simple but powerful grouping and both text and gui result interfaces. In order to use it you need plt scheme. Then you can grab the unit test suite from planet:
(require (planet "test.ss" ("schematics" "schemeunit.plt" 1 2)))

The code here shows testing only the any/text type and all relevant auxilary functions, using this datatatype. Looking at the code, it could be turned into a macro, so we could generate type tests automatically, rather than writing this long code..