J3/00-167

Date:        5 May 2000
To:          J3
From:        R. Maine
Subject:     F90 interp 13 and the f2k IMPORT

Paper J3/99-197R1 presented several options for dealing with
the interface body scoping problem.  A straw vote favored
addition of an IMPORT statement.  However, it occurs to me
that there may be a simpler and less confusing solution than
any of those presented in that paper.

I suggest that we review and reconsider f90 interpretation 13.
I recall this interp as one that was being debated at my very
first J3 meeting (meeting 119).  It was very controversial
from the start.  The f90 standard, as originally published,
unambiguously said that host association worked with interface
bodies just like it worked elsewhere.  There were even
examples illustrating this.  Interp 13 said that this was an
error and revised f90 to change it.  There was a bit of
controversy over whether or not to make interface bodies an
exception to the host association rules.  There was perhaps
even more controversy over whether it was acceptable to change
something that was unambiguous and consistent in the standard
as published.

I don't propose to revisit the issue as an f95 interp.  I didn't
like retroactive revision when interp 13 was being debated, and
I still don't like it.  I figure that interps are to correct
actual errors rather than to do retroactive design.

But we could still change this in f2k.

It seems to me that the whole business of adding an IMPORT
statement is just a workaround to a problem introduced by
interp 13.  I propose fixing the problem instead of working
around it.

We already have users regularly being confused by the fact
that host association doesn't go into interface bodies.
Check comp.lang.fortran and comp-fortran-90 archives for
regular reoccurances of this confusion.  The IMPORT statement
allows you to specify that a specific list of names is
imported to an interface body, just like they would have
been in the first place without interp 13.  Adding a whole
new statement to the language to hack around a bug strikes
me as questionable.

Wouldn't it be much simpler to just have f2k allow host
association into interface bodies just like in all other
scping units?  Effectively reverse interp 13 for f2k.  There
then would be no need to train users about how host
association works differently for interface bodies.  And there
would be no need for an IMPORT statement to fix it up in the
cases where you really need it.

Note that the stated goals of interp 13 are defeated by the
IMPORT statement anyway.  With the IMPORT statement, you are
no longer be guaranteed that the same text of an interface
body will be interpreted identically independent of the host.

The advantages of reversing interp 13 seem obvious to me.
The language would be simpler in that there would be no
special case for host association into interface bodies
and there would be no need for a new statement.

The disadvantages seem small in comparison.

1. There would be a small incompatability with f90/f95 for
   rare cases of badly-written code.  If an interface body
   depended on default implicit typing and was in a host with
   non-default implicit typing, then the interface body
   might have a different interpretation.  There's an
   excellent chance that this is the interpretation a
   user would have expected anyway.

   This would be appropriate to list in the introduction
   under incompatabilities.

2. There would be more ways to write interface bodies that
   had different interpretations in different hosts, and such
   dependency wouldn't be as immediately obvious as it is
   with an IMPORT statement.

   But I claim that this is no different from lots of other things
   in Fortran, that a well-written interface body won't run into
   this, and that it will cause less confusion than is quite
   clearly caused by the current state of affairs.

   You don't *HAVE* to have such dependence, except for the cases
   where you'd need an IMPORT statement and thus be so dependent
   anyway.

   Indeed, this is very much like the (true) statement that a
   subprogram might have a different interpretation if you move
   it to a different host and it depended on host association.

I believe that experience has shown interp 13 to be a mistake.
I propose that we fix that mistake in f2k rather than patching
around it.