Incorrect array strides of sliced array passed as a Fortran argument

[sbrdar]

What happened?

On the nvfortran 22.11 the compiler repacks sliced arrays passed as arguments to a contiguous temporary array. This return wrong strides of the sliced array.

What are the steps to reproduce the bug?

On nvfortran 22.11 (and possibly other NVHPC compilers since they repack array by default) the following fckit strides computation is incorrect:

use fckit_array_module, only : array_strides
integer(c_int) :: existing_data(8, 2, 3, 4)
print *, array_strides( existing_data(1, :, 2, :)) )

This returns [1, 2] instead of expected [8, 48].

If the flag -Mnotarget_temps is passed, the nvfortran compiler will not repack the temporaries and then we get the correct results.

Version

develop

Platform (OS and architecture)

nvfortran 22.11

Relevant log output

Accompanying data

No response

Organisation

No response

[wdeconinck]

From NVIDIA HPC Compilers Reference Guide :

-M[no]target_temps
instructs the compiler to enable [disable] using temporaries when passing an array for a callee assumed-shape variable with the target attribute.

So this only works with arrays containing the target attribute, which is what is still missing in the array_strides subroutine.

From the NAG documentation:

-Oassumed=shape
Optimises assumed-shape array dummy arguments according to the value of shape, which must be one of
always_contig
Optimised for contiguous actual arguments. If the actual argument is not contiguous a runtime error will occur (the compiler is not standard-conforming under this option).
contig
Optimised for contiguous actual arguments; if the actual argument is not contiguous (i.e. it is an array section) a contiguous local copy is made. This may speed up array section accessing if a sufficiently large number of array element or array operations is performed (i.e. if the cost of making the local copy is less than the overhead of discontiguous array accesses), but usually makes such accesses slower. Note that this option does not affect dummy arguments with the TARGET attribute; these are always accessed via the dope vector.
section
Optimised for low-moderate accesses to array section (discontiguous) actual arguments. This is the default.
Note that CHARACTER arrays are not affected by these options.

The section on contig states also that even with this option, variables with target attribute are not repacked. Then that could mean that the previously reported issue ecmwf/atlas#213 could also possibly be solved here by adding the target attribute.

[wdeconinck]

Tested with nvhpc 25.9, the flag "-Mnotarget_temps" is available on Linux_x86_64 as on ECMWF's AC, but not on Linux_aarch64 as on ECMWF's AG with GH200.

[wdeconinck]

This is being worked on in #74