On Mon, 8 Mar 2004 09:51:04 -0800, Richard Maine <[log in to unmask]> wrote:
> At one time, it was actually on J3's approved list as a feature to be
> added. It got dropped because of its difficulty; it isn't easy.
I looked through the standard documents archive. I think you're
refering to number 035 of requirements for F2000 (ca. 1995-1996).
> Reals aren't the only issue, though. In some ways, reals are simpler
> to handle than integers. Even scientific codes tend to involve many,
> many integers, probably more than reals though I haven't bothered to
> take actual statistics. Heck, for every array, you probably have one
> or more integers involved in declaring its bounds for a start. (I'm
> counting an array as a single case; if you count each array element
> separately, then reals likely dominate by a lot because there are
> often large arrays of reals). If you had to put a kind number on
> every integer, it would probably make most codes more inscrutable than
> if you had to put a kind number on every real.
>
> Many large scientific codes do exceed the limits of 32-bit integers,
> so yes, it is a significant issue.
Well, I was thinking of just skipping integers and dealing with
only real types, but.. Maybe a question that has to be asked first
is if the defaults will change in the future?
By the time F2003 is as widely adopted as F95 is today, 64-bit
integers will be the more natural one on most hardware (that's PCs).
Will Fortran have default 64-bit integers then? If it will, then
precision of default real will be made better just by waiting,
and it may be the best solution.
Maybe it's more productive to think of ways to ensure that the
de-facto default change to 64-bits?
--
Yasuki Arasaki
[log in to unmask]
|