Re: [Caml-list] Patterns that evaluate

[skaller <skaller@users.sourceforge.net>]

On Wed, 2007-02-14 at 16:10 -0500, Jacques Carette wrote:
> Nathaniel Gray wrote:

> > I guess I'm not seeing it.  How did you expect it to work?  Is this
> > what you mean:
> >
> > ... | ({st = Some v}, _) when v = e -> e2
> Yes. that is what I meant.
> 
> > Is there some functionality that you're looking for that when clauses
> > don't provide?
> No, there is not.  I was just (mistakenly) assuming that the "more 
> pleasant" syntax for this ``worked''.

It is a common wish, but has many problems IMHO.

First, it isn't very general. When clause has form:

	  | ... x ... when P(x) ->

which can express a lot more than merely one of the three
equality operators. Indeed, P doesn't even have to be a
known function -- it can be a function value.

Second, the translation:

	| .. e .. ->   to | .. x .. when x = e -> 

without extra lexical marks distinguishing
pattern variables from local variables, and more generally
patterns from expressions, would be extremely fragile:

	let h = 1 in 
	match x with
	| h :: t -> ...

The meaning of this matching appears to be to match
a list starting with 1, binding t to the tail .. but
it all depends on whether t is a symbol in the context
of this fragment. The meaning would change if t were
not in the context and a programmer introduced it:

(* no t here *)
....
let ...

     | h :: t ->

Now the programmer adds a new global variable

let t = [2]

...


and it changes the meaning of the implementation of
a function which one would expect was abstracted and localised.

Thirdly it leaves open the issue of non-linear patterns:

	| x , x -> ...

The first x binds to the first component of a tuple ..
what does the second x do?? Easily solved with a when
clause:

	| x, y when x = y -> ...

Fourth, when clauses have a restriction that they do not
allow introduction of extra bindings. I often need that:

	| C  x | D with x = 1 -> ... x ....

where I'm introducing a default for one branch. 
I also often match nasty patterns in many places and really
wish I could name them:

	pattern p(x) = C x
	match e with | p(x) -> ...

doesn't provide active patterns, but it would be quite useful.
This can be done with a match composition already too:

	let p e = 
		try match e with C x -> x 
		with _ -> raise MatchFailure
	..
	try let x = p e in ..with MatchFailure -> (* next case *)

but it is messy (hard to compose).

Jacques said:

> Is there some functionality that you're looking for that when clauses
> > don't provide?
> No, there is not.

and that's basically wrong. He's not telling the truth here :-> 
He IS looking for something more general, specifically

	| f a -> ..

where f is pattern match variable. Ocaml currently requires f to be
a constructor, that is, a constant. Any notion of advanced
pattern matching would have to allow match variables to bind
to constructors. 

IF we're going to extend pattern matching it should do this:

http://www-staff.it.uts.edu.au/~cbj/Publications/pattern_calculus.ps

IMHO of course.. :)

-- 
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net