From: Joel VanderWerf <vjoel@...>
Date: 2001-09-02T12:10:56+09:00
Subject: [ruby-talk:20687] Re: Iterating by links

David Alan Black wrote:
> What I'm finding is that it's hard to automate because there are
> different things that could cause the exception.  For instance, in

David,

Semantic issues aside, the implementation using exceptions is problematic. 
One possible fix: in the "rescue NameError" clause, compare the exception's 
backtrace depth with the current stack depth (i.e., caller.size). 
Apparently, these differ by 4 if and only if the exception was generated 
within the loop, so we can re-raise just those exceptions caused by user 
code.

However, using a constant in this way makes me itch, so I considered going 
back to the respond_to? approach, which is at worst about 20% slower than 
using exceptions to test implicitly.

But then there are the semantic questions: What terminates the enumeration? 
A nil return value from send? An object that doesn't respond to the 
message? Either? And are terminating objects included in the enumeration?

My current feeling (I've changed 180 degrees on this) is that an object 
that doesn't respond is actually an error condition, not a termination 
condition. This means that my tree example simply won't work as is (but see 
below). 

There are programming errors that will be ignored unless we do it this way. 
(Perhaps you have several classes of objects, and you forgot to define a 
'next' attribute for one of them, or you simply linked the wrong object 
into your list.)

Also, nil (or false) terminates the enumeration without being included in 
it, which is the classic linked list semantics.

Here are two ways to iterate over a tree branch:

  tree = [[0, [1, 2]], 3, [4, 5, [6, 7, 8]], 9]

  at_random = proc { |x| x.kind_of?(Array) && x.at(rand(x.size)) }

  for node in tree.by at_random
    p node
  end

  class Object
    def at_random
      nil
    end
  end
  class Array
    def at_random
      at(rand(size))
    end
  end

  for node in tree.by :at_random
    p node
  end

The latter is more elegant and probably more efficient, but the former is 
nice if you just want to do it on the fly.

I'll release the updated version on the web, along with some 
other enumerable stuff, later this weekend.

BTW, is there some reason why LinkedListDelegator shouldn't be hidden away 
inside Enumerable's namespace, rather than in the global namespace?

Thanks for your helpful comments!

Joel