some suggestions from Richard Weed, author of F90 API

[edhartnett]

Here's an email I got from Richard Weed, who wrote the F90 API many years ago. Some of these suggestions seem promising and I will take a look when I have cleared some of the low-hanging fruit away.

Hi Ed.

I'm still at this address. Unfortunately
I don't have much (any really) time to work
on netCDF now However, if I did, here are the things
I would probably look at.

1. My current implementation really needs to
   be refactored to reduce some of the (in hindsight)
   extraneous code I implemented (almost 13 years ago
   now) when the F2003 C-Interop facility was still
   in its infancy. An example is the two modules that
   contain the C interop interfaces (module netdcd_nc_interfaces.f90
   and module_netcdf4_interfaces.f90) etc. In retrospect
   a better design option would have been to embed the interfaces
   in the routines that actually use them. This is what HDF5 does.
   This would allow you to get rid of those modules and not
   have to include their resulting .mod files in your include
   directories. My original decision to implement the interfaces
   this way reflected my coding practices back then. My thinking
   was that this approach mimicked the equivalent C header files.
   Removing then would eliminate the need for users to make sure
   they had the correct paths to the NetCDF include directories
   in their build scripts.

2. The most recent edition of the Fortran standard (Fortran 2018
   released in December) defines a lot of enhancements to the C-interop
   facility that offers the promise of greatly reducing the amount
   of code you currently have to support all the multiple data types
   and precisions along with the current preprocessor code you
   have to support all the array ranks. In particular, three new facilities
   for passing arrays of any type, any rank as well as support for
   assumed shape Fortran dummy arguments which are passed as structures
   that contain more than just the address in memory ( ie. bound,
   rank, and kind information) so they are not directly interoperable
   with C. See the discussion of assumed type, assumed rank, and the
   C descriptors (supplied by a mandated C header file named
   ISO_Fortran_binding.h) in the attached PDF. Unfortunately,
   only the most recent versions of the Intel compiler (18 and 19)
   and probably the Cray compiler have full support for the new
   2018 features. gcc/gfortran has them implemented in their
   current development trunk (9.0) BUT I don't have any hope
   they will actually be usable (bug free) for a couple more
   years. The new version of the flang/LLVM compiler front end being
   developed by nVidia (called f18) will support them but it will
   probably be the end of the year before you can
   actually compile anything with it.

3. I haven't kept up with any new functions you have in the
   C interfaces so I can't comment on the best way to support
   them. If I remember, the biggest missing capability in
   the Fortran interfaces was the ability to support arrays
   of strings (specifically arrays of variable length strings)
   That is doable with F2008/F2018 but is not directly supported
   by the standard Fortran 90 interfaces and probably
   can't be made backwards compatible to the Fortran 77
   interfaces so you would need to make sure users understood they would
   need a recent compiler to support them and make sure you had sufficient
   documentation on how to use them. Currently, the best
   way to implement arrays of variable length strings is to embed
   a F2003/2008 deferred length string in a derived type. I have
   a whole package of code I use to do this for other projects I
   can send you. The issue then becomes passing the derived type
   (structure) to C. An efficient implementation that makes
   thing as simple as possible for users will probably require
   some operator overloading for assignments etc and some
   new functions you would have to document.

Sorry I can't take a more active roll now (see
me in 2 or 3 years when I (hopefully)
retire). However, I will try to answer any Fortran programming
or implementation questions you have.

RW