From: Mark Hubbart <discord@...>
Date: 2004-04-20T07:33:07+09:00
Subject: proposal: call_up() for use in redefined methods

Hi all,

A common ruby idiom for redefining a method in a class is to alias it 
to give it a new name before defining the replacement:

   # allow integer indices to wrap.
   #   "test"[6]  #=>"e"
   class String
     alias :old_brackets :[]
     def [](idx, num=nil)
       idx = ((idx % size) + idx) % size if idx.kind_of? Fixnum
       num ? old_brackets(idx, num) : old_brackets(idx)
     end
   end

However, this breaks if it is done a second time, with the same alias:

   # allow ranges to wrap around
   #   "abcd"[-2..1]  #=>"cdab"
   class String
     alias :old_brackets :[]
     def [](idx, num=nil)
       if Range === idx and idx.first < 0 and idx.end > 0
         end_range = idx.exclude_end?? (0...idx.end) : (0..idx.end)
         self[idx.first..-1] + self[end_range]
       else
         num ? old_brackets(idx, num) : old_brackets(idx)
       end
     end
   end

then in irb:

   irb(main):004:0>>> "abcd"[3]
   (eval):4:in `old_brackets': stack level too deep (SystemStackError)
           from (eval):5:in `old_brackets'
           from (eval):5:in `old_brackets'
   <snip long stack error>

This can be a problem. If those two definitions occurred in two 
separate libraries, then the maintainers of both libraries would have 
to get together (once the problem is identified) and modify their code 
just so the libraries can be used together. It could discourage people 
from using what is (IMHO) one of Ruby's greatest strengths... the 
ability to easily redefine methods.

One solution comes from the common OO method/construct: super(). 
Calling super() lets you call the parent class' version of a particular 
function; it's smart, and knows to call the same function that you are 
currently in. Since most OO languages (that I've worked with) have 
non-redefinable classes, super() is enough. But when you redefine a 
method in Ruby, there's no way of getting to the previously defined 
method. I propose a new method/construct, tentatively named call_up(), 
which would give access to the older versions of a method; and a 
companion method redefine() which stores the method in a stack before 
undefing it.

Here's a partial example of a hacked solution. Additional coding would 
be needed to support the same thing in modules and singleton classes, I 
but it shows a rough idea of the way it would work.

   class Object
     def redefine(sym)
       if implements? sym
         @@__redefined_methods[sym] ||= []
         @@__redefined_methods[sym].push method(sym)
         undef sym
       else raise ArgumentError, "not a valid method"
       end
     end

     def call_up(*args)
       # Deep Magic Goes Here:
       # - detects the calling method
       # - detects if it is in the __redefined_methods stack
       # - determines which method version was defined immediately 
previous
       # - calls that method with *args
     end
   end

Each class (and singleton object) would keep track of the explicitly 
redefined methods, and allow a super()ish way of calling the next 
method up. This could allow the above method redefinition code to look 
like this:

   # allow integer indices to wrap.
   #   "test"[6]  #=>"e"
   class String
     redefine :[]
     def [](idx, num=nil)
       idx = ((idx % size) + idx) % size if idx.kind_of? Fixnum
       num ? call_up(idx, num) : call_up(idx)
     end
   end

   # allow ranges to wrap around
   #   "abcd"[-2..1]  #=>"cdab"
   class String
     redefine :[]
     def [](idx, num=nil)
       if Range === idx and idx.first < 0 and idx.end >= 0
         end_range = idx.exclude_end?? (0...idx.end) : (0..idx.end)
         self[idx.first..-1] + self[end_range]
       else
         num ? call_up(idx, num) : call_up(idx)
       end
     end
   end

If we could do this, it would eliminate any worries about conflicting 
method names, and would free Ruby programmers to redefine methods 
whenever it makes sense. I don't think it would introduce any real 
added complexity, and it would standardize how to redefine a method and 
call the original.

For an even more real-world example, consider all the different numeric 
types defined in the stdlib math libraries. If you want to create a new 
Numeric, and want Math.sqrt() to support it, you have to redefine 
Math.sqrt. But if it just happens that someone else did the same thing, 
in the same way, then everything will go horribly wrong the moment you 
use Math.sqrt() after require()ing both files.

Comments? Questions? Flames? :)

--Mark