John Reid wrote:
> > Given that copying is not the only (or even the most common or
> > useful) reason that aliasing is traditionally disallowed, that is
> > a non-sequitur decision. Optimization is far more important
> > a reason to disallow aliasing. It remains important even when
> > arguments have TARGET attribute.
>
> This is your opinion, but it is not everyone's view. The committee
> was faced with protests from people using pointers and targets that
> were not getting the results they expected.
I side with J. Giles here, being an HPC programmer myself. I think Fortran
made array pointers soo inefficient by basically throwing away most of the
restrictions otherwise applicable. I love them as objects and they are soo
useful, but I have to avoid using them so as not to sacrifice efficiency. It
is rather painful...TR15881 convinced me not to give up on Fortran 95. Now I
have to start avoiding TARGET whenever possible. But do understand why people
were complaining--the standard is complicated in this area so what you expect
may not be what you get :(
Of course, all IMHO.
Aleksandar
|