J3/06-241

Date: 31 July 2006
To: J3
From: Bill Long and John Reid
Subject: Issues 20, 23 : co-indexed variable
Reference: J3/06-007


Discussion of Issues 20 (page 117) and 23 (page 157)

Remote allocation has never been part of CAF and we believe that there
is agreement about keeping this prohibition.

The disagreement is over the rules for
     Z[P] = Z
when Z is of a derived type with an allocatable component.

The subgroup wishes to allow this provided the shapes conform. For
example, the programmer may arrange that a team of images allocate
the component to be of a single shape, while the rest make no use of
it.

The editor claims this is inherently unsafe. Why? Can't the user
inquire about the remote bounds, as in:

IF(SHAPE(Z[P]%allocatable)==SHAPE(Z)) THEN
   Z[P]=Z
ELSE
   STOP "Mismatch in shape"
END IF

After all, we only added automatic reallocation for allocatable arrays
in F2003. Before that we had the same thing for allocatable
arrays---the user had to make sure the shapes agreed in intrinsic
assignment.

Concerning the fact that this gives different meaning to intrinsic
assignment when a co-array is involved, we don't agree this is
confusing. The user must be thinking about these issues---parallel
programming is more complicated. Remote assignments are more
restricted than local ones. It should be the first sentence in a
textbook.

A straw vote is needed to decide this.

Edits to J3/06-007

If intrinsic assignment for co-indexed objects with allocatable
components is retained, then [157:10-11] needs to be deleted.