(j3.2006) (SC22WG5.3615) Preparing for the Tokyo meeting

[Lawrie Schonfelder]

I am fed up with making comments and then being bombarded with replies that answer opinions that I
do not hold and have not expressed.

I am not opposed to co-arrays. I believe they provide a rather better mechanism for programming code
that has a fighting chance of being implemented efficiently and effectively on a number of common
current architectures. I believe they should be standardised.

but

I am not convinced they are the only, or even best way of achieving this. No, I do not know of a
better way but I am aware that there have been a number of break threw moments of great ideas over
the past decade or so that have been and gone. I do not therefore want the core Fortran language to
be saddled with co-arrays as an integral part of the language, irremovable for at least 20 and
perhaps more years.

I would like co-arrays to be standardised as an optional standard. I would also like all the
necessary support for MPI to also be standardised, probably as an option.

This is not a technical argument. It is not about a better "go faster stripe". It is that this go
faster stripe should be standardised as an option which if it is superseded by a better one in ten
years time is not stuck cluttering a language already cluttered with legacy structures that would be
better not there. My view is one of language design principle. The core language should be aimed at
the effective expression of algorithms independent of architecture. Additional facilities that allow
certain algorithms to be expressed more efficiently on particular architectures should be optional.
I believe co-arrays are just such an addition.

I do not accept co-arrays are in any way analogous to COMPLEX. COMPLEX is a data type that has real
world meaning and its inclusion within the language supports the programming of a major set of
problems regardless of architecture. Co-arrays are in principle an artefact of a particular way of
mapping problems onto a specific type of machine architecture. They have no real world physical or
mathematical model correspondence. Hence my opposition to their inclusion in core Fortran.


--
Lawrie Schonfelder
Wirral, UK