W3C home > Mailing lists > Public > www-dom@w3.org > October to December 1998

Re: Walking the DOM (was: XML APIs)

From: Stephen R. Savitzky <steve@rsv.ricoh.com>
Date: 03 Nov 1998 13:45:22 -0800
To: John Cowan <cowan@locke.ccil.org>
Cc: DOM List <www-dom@w3.org>
Message-ID: <qck91cqwlp.fsf@gelion.crc.ricoh.com>
John Cowan <cowan@locke.ccil.org> writes:

> Stephen R. Savitzky wrote:
> > but I think you may have missed something.  An
> > iterator, the state of which consists of a reference to the ``current
> > node'', and all of whose methods are inlineable, can be compiled by a good
> > optimizing compiler into the exact equivalent of your algorithm.
> Can you sketch this for me?  I don't see how the recursive depth-search
> can be optimized away, since the iterative version depends on the
> existence of the parent pointer, which the original algorithm did not
> even mention.  The breadth-search is obviously tail recursive and
> no fundamental problem.

I was really referring to the iterator given in the next paragraph:

class depthFirst {
  Node node;
  Node toNext() {
    if (node.firstChild != null) {
	node = node.firstChild; return node;
    } else {
	while (node.nextSibling == null) {
	    node = node.parent;
	    if (node == null) return null; 
	node = node.nextSibling;
	return node;

If you declare toNext as "inline" and an instance of depthFirst as local,
any C++ compiler would get roughly the equivalent of your loop.

But in fact the simple recursive depth-first search could probably be made
iterative if the compiler knew about the identity
  node.firstChild.parentNode = node
(Which as you point out is not an identity except in a pure funtional-
programming environment).

> > Similarly,
> > your algorithm cannot be used as the basis for a live NodeList.
> No, indeed.  It is used in my DOMParser, which walks a (presumed static)
> DOM in order to generate SAX events from the nodes.

AHA!  If it's ``presumed static'', then your NodeList's aren't live!
According to the DOM spec you can't make that presumption, so in fact _your_
implementation isn't complient to the spec any more than mine is, which is
_precisely_ my point. 

I would like to see a spec that, essentially, parametrizes these
presumptions in order to permit a straightforward, efficient

> > I believe
> > it is also essential that DOM implementations make structural _modification_
> > methods fast, which is also incompatible with the kind of non-locality of
> > effects imposed by live NodeLists and ``robust'' (in your definition)
> > iterators.
> I hope that it is *not* utterly incompatible, but of course "fast"
> is a loosely defined term; if you mean O(1), you are probably
> correct.

I DO mean O(1).  Without live nodelists and all the other handholding, an
O(1) implementation of _all_ the structural methods in the Node interface
is not only possible, it's the _most obvious_ implementation.  It's what any
sane programmer _would_ implement.  

My whole point is that _with_ live NodeLists, the simple, sane, obvious,
lowest-common-denominator implementation becomes impossible. 

At a rather fundamental level, the interfaces themselves specify an
implementation.  You could take the IDL interfaces for Node and its
subclasses, add pre- and post-conditions for the methods, run them through a
code generator, and get an implementation.

The whole problem with NodeList as specified is _precisely_ that it breaks
the encapsulation of Node by rendering this straightforward implementation
invalid.  This is the sense in which the specification is incorrect.
Actually, there's another sense: I think that it would be difficult to
express the ``live NodeList'' requirement in a formal specification
language.  It may be _impossible_ in an OO modeling language like UML.

There's a well-known principle in user-interface design called, if I
remember correctly, the principle of ``least surprise.''  A program should
not behave in ways its user would find surprising, given its user interface.
The user of an API, for example, is an application programmer.

>                     I concede that I don't yet see how best to
> do it, which is (one reason) why I have not published a DOM
> implementation.

My point is that the problem is not with your implementation, but with the
specification.  Something that is impossible to implement efficiently is not
going to get used very much.  People like you and me are going to implement
the interfaces in the obvious way, and put disclaimers in the comments to
the effect that ``this application implements the DOM interfaces, but not
all of the specified behavior.''

Then if somebody takes our application and replaces our hacked DOM class
library with a ``real'' one, the performance will go to blazes and we'll get
the blame.

 Stephen R. Savitzky   Chief Software Scientist, Ricoh Silicon Valley, Inc., 
<steve@rsv.ricoh.com>                            California Research Center
 voice: 650.496.5710   fax: 650.854.8740    URL: http://rsv.ricoh.com/~steve/
  home: <steve@starport.com> URL: http://www.starport.com/people/steve/
Received on Tuesday, 3 November 1998 16:40:10 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Friday, 22 June 2012 06:13:46 GMT