(j3.2006) a question on c_f_pointer call

[Aleksandar Donev]

On 08/27/10 11:57, Bill Long wrote:
> For many implementations (excluding NAG, evidently) the needed 
> information is included in the descriptor.  Until we introduced 
> C_F_POINTER and C_LOC, that worked fine. 
Which descriptor?!? There is the pointer/allocatable array descriptor 
that was originally filled in by ALLOCATE. Sure, setting a field there 
is easy. But then, this field has to be *consistently* copied at every 
pointer assignment and at every procedure call. If you pass the array 
created by allocate to a TARGET dummy, someone can point a pointer to 
the dummy. You can pass sections, and someone can pointer a pointer to 
that section. They can do pointer rank remapping, and go through several 
pointers of different rank, and then back. As long as in the end one 
gets a pointer that points to all of the elements of the original target 
created by allocate, your flag in the descriptor has to be set to 
"eligible for deallocation". How you can do this without any 
calculations at pointer assignments or procedure calls is beyond me. 
Just copying the flag does not work---you have to look at whether the 
pointer is "associated with the whole of an object that was created by 
allocation...in array element order". This has some overhead.

If you asked me, I would not have required error reporting, so that 
different implementations are all valid and you can argue internally in 
your team about which overhead is bigger or more frequent than other 
overheads. But to claim that c_f_pointer changed this in some 
fundamental way still seems wrong to me.

Best,
Aleks

-- 
Aleksandar Donev, Assistant Professor of Mathematics
Courant Institute of Mathematical Sciences
Office: 909 Warren Weaver Hall, New York University
E-mail: donev at courant.nyu.edu
Phone: (212) 992-7315; Fax: (212) 995-4121
Mailing address: 251 Mercer St, New York, NY 10012
Web: http://cims.nyu.edu/~donev