[GAP Forum] Forum discussion on packages and global function names

[Steve Linton]

On Thu, 03 May 2007 08:21:31 -0500
Yevgen Muntyan <muntyan at tamu.edu> wrote:


> Why should Attributes be preferred to Operations? Does it make something
> faster? We can have operation with the same name declared and defined
> for totally different objects, so the only difference between operation and
> attribute seems to be that attribute value is stored. Or is there something
> else?
> 

The reason is generality. If you declare something as an Attribute and never
call it for any Attribute-storing object, then nothing is lost. But if you
declare it to be an Operation, and someone comes along later and wants to use
it for Attribute-storing objects, they will not be able to take advantage of
the automatic storing and reuse of results.

If you have an Operation which takes only one argument and is not defined to
have any side-effects, then it should be declared as an Attribute unless you
have a strong reason not too.
> 
> [snip]
> > I should say that there is one awkward exception to this generally good
> > performance, namely when one of the arguments is a plain list with mutable
> > entries. In this case determining its type must be repeated at every operation
> > call, and can be expensive. We have some longer term plans to address this,
> > but thy will involve significant changes and will take a while to work through.
> >   
> Do you mean that it takes much time to distinguish between
> 
> InstallMethod(Foobar, [IsList], ...)
> InstallMethod(Foobar, [IsList and IsSomeCollection], ...)
> 
> or any operation which has a method installed for a list is affected?
> 

It has nothing to do with the Operation or the methods installed for it.
Whenever an Operation is called, the types of each of the arguments is
determined. Most of the time this is constant time, or, if significant work is
involved, the results of the work are stored so that it will only need to be
done once. For lists with mutable entries, however (not just mutable lists,
but lists at least one of whose entries is a mutable object), this information
cannot safely be stored, and so must be checked each time, and this can be
expensive.

	Steve
-- 
Steve Linton	School of Computer Science  &
      Centre for Interdisciplinary Research in Computational Algebra
	     University of St Andrews 	 Tel   +44 (1334) 463269
http://www.cs.st-and.ac.uk/~sal 	 Fax   +44 (1334) 463278