Re: [Patch, Fortran] Support allocatable *scalar* coarrays

[Tobias Burnus <burnus at net-b dot de>]

Mikael Morin wrote:
> let me understand one thing about coarray scalars: despite their name, they
> are arrays, right?

Yes and no. In terms of the language, they are scalars - but they have a 
codimension, e.g.
    integer, save :: A[4:6, 7:*]
is a scalar variable on each image, but it has a coarank of 2 with 
lcobound(A) == [4, 7] and ucobound(A, dim=1) == 7. (The value of 
cobound(A, dim=2) depends on the number of images, it's >= 7 in this 
example.)

In terms of gfortran, nonallocatable coarrays are normal scalars - with 
a lang-specific node attached to them, which contains the cobounds, i.e.,
  GFC_ARRAY_TYPE_P (type) = 1;
  GFC_TYPE_ARRAY_CORANK (type) = as->corank;
with
  GFC_TYPE_ARRAY_LBOUND (type, dim)
containing the trees for dim = (rank + 1) ... (rank + corank).

The same scheme is used for assumed-type coarrays:
  subroutine sub(B, n)
     integer :: B(:)[5:7, n:*]

Note that here that contrary to the dimension, the codimension is not 
":" (i.e. assumed shape) but that it is assumed-size.


For allocatable (scalar) coarrays, one has:
   integer, allocatable :: B[:, :]  ! Note: The coshape is deferred
   ...
   allocate (B[2:3, 5:*])

Again, one has the actual data and the cobounds. For that case, I have 
decided to store the information in the array descriptor of rank == 0 
and dim[0 ... corank-1] for the bounds. Thus, "desc->data" contains the 
scalar but the variable itself is a descriptor (GFC_DESCRIPTOR_TYPE_P). 
The corank is not stored in the descriptor, but as one knows the number 
of codimensions (an explicit interface is required for allocatable 
coarray dummies), one knows the corank.

> Then when you do in gfc_conv_array_ref:
>
> +      if (GFC_DESCRIPTOR_TYPE_P (TREE_TYPE (se->expr)))
> +       se->expr = build_fold_indirect_ref (gfc_conv_array_data (se->expr));
> [...]
>         return;
>
> you are returning scalar[1] instead of scalar (== scalar[this_image()]) or
> scalar[whatever_image_selector], aren't you?

Well, the current implementation supports effectively only a single 
image - for -fcoarray=single on purpose and for -fcoarray=lib because it 
has not yet been implemented.

Later, one has to add some function call for "scalar[<image_numer>]" 
while "scalar" itself is the local variable and can be handled as above. 
The expression of "scalar" ends up having expr->ref->type == REF_ARRAY 
with dimen_type == DIMEN_THIS_IMAGE. That way one can distinguish a 
reference to the local coarray and to a remote coarray (coindexed 
variable); note that "coarray[this_image()]" also counts as 
remote/coindexed.

> Sorry for the delay; it seems that the more it goes, the more you are the only
> one who can maintain coarray stuff. :-(

Well, Daniel Carrera develops into an trans*.c, allocate, 
libgfortran/caf/ expert :-)

Tobias

PS: I should document somewhere how coarrays are implemented internally.