[3/3] Check array contiguity for OpenACC/Fortran

Hi,

This patch tightens up error checking for array references used in
OpenACC clauses such that they must now be contiguous. I believe this
matches up to the spec (as of 2.6). I've tried to make it so an error
only triggers if the compiler is sure that a given array expression must
be non-contiguous at compile time, although this errs on the side
of potentially allowing the user to shoot themselves in the foot at
runtime: at least one existing test in the testsuite appears to expect
that behaviour.

This hopefully addresses Tobias's concern in:

  https://gcc.gnu.org/ml/gcc-patches/2019-12/msg01439.html

In particular, with the error checking, we no longer get an ICE for the
example given in that message. The new check also catches a test case
that appears to have been relying on undefined behaviour.

Tested with offloading to NVPTX. OK for trunk?

Thanks,

Julian

ChangeLog

2020-01-04  Julian Brown  <julian@codesourcery.com>

	PR fortran/93025

	gcc/fortran/
	* openmp.c (resolve_omp_clauses): Check array references for contiguity.

	gcc/testsuite/
	* gfortran.dg/goacc/mapping-tests-2.f90: New test.
	* gfortran.dg/goacc/subarrays.f95: Expect rejection of non-contiguous
	array.
---
 gcc/fortran/openmp.c                          | 29 +++++++++++++----
 .../gfortran.dg/goacc/mapping-tests-2.f90     | 32 +++++++++++++++++++
 gcc/testsuite/gfortran.dg/goacc/subarrays.f95 |  2 +-
 3 files changed, 55 insertions(+), 8 deletions(-)
 create mode 100644 gcc/testsuite/gfortran.dg/goacc/mapping-tests-2.f90

Hi Julian, hi Thomas,

in terms of the check, it looks fine to me – but I am not sure about the 
spec.

At least the following test case seems to work fine:

integer :: A(10,10), out(12)
A = reshape([(i, i=0,100)], shape(A))
!$omp target map(A(3:6,3:5), out)
!$acc parallel copy(A(3:6,3:5), out)
   out = reshape(A(3:6,3:5), shape(out))
   A(3:6,3:5) = 4
!$acc end parallel
!$omp end target
print '(4i3)', out
print '(10i3)', A
end

The reason that it works is that the stride is included in the length calculation:
#pragma omp target num_teams(1) thread_limit(0) map(tofrom:MEM[(c_char *)_6] [len: _5])

Has for the section (with A(3:6,3:5) -> parm(1:4,1:3)):
               parm.1.dim[0].lbound = 1;
               parm.1.dim[0].ubound = 4;
               parm.1.dim[0].stride = 1;
               parm.1.dim[1].lbound = 1;
               parm.1.dim[1].ubound = 3;
               parm.1.dim[1].stride = 10;
And instead of doing '(4-1+1) * (3-1+1)' = 12 (i.e. multiplying the extends),
the code does: 'stride * (3-1+1)' = 30.

Cheers,

Tobias

PS: It also works if one puts the stride on the ptr, i.e.:

integer,target :: A(10,10), out(12)
integer, pointer :: ptr(:,:)
A = reshape([(i, i=0,100)], shape(A))
ptr => A(3:6,3:5)
!$omp target map(ptr, out)
!$acc parallel copy(ptr, out)
   out = reshape(ptr, shape(out))
   ptr = 4
!$acc end parallel
!$omp end target
print '(4i3)', out
print '(10i3)', A
end

On 1/4/20 3:25 AM, Julian Brown wrote:
> This patch tightens up error checking for array references used in
> OpenACC clauses such that they must now be contiguous. I believe this
> matches up to the spec (as of 2.6). I've tried to make it so an error
> only triggers if the compiler is sure that a given array expression must
> be non-contiguous at compile time, although this errs on the side
> of potentially allowing the user to shoot themselves in the foot at
> runtime: at least one existing test in the testsuite appears to expect
> that behaviour.
> […]

(CC: fortran@ as it relates to Fortran.)

Hi all,

On 1/7/20 12:16 PM, Tobias Burnus wrote:
> in terms of the check, it looks fine to me – but I am not sure about 
> the spec.

* [OpenACC] Actually, I simply missed the bit (here: OpenACC 3; OpenACC 
2.6 is same): “Any array or subarray in a data clause, including Fortran 
array pointers, must be a contiguous block of memory, except for dynamic 
multidimensional C arrays.” (2.7.1 under Restrictions).

* OpenMP (quoting TR8): “If a list item is an array section, it must 
specify contiguous storage.” (2.22.7.1 map Clause under Restrictions). 
However, that one seems to miss the case: “non_cont_ptr => A(::2); !$omp 
... map(non_cont_ptr)” as non_cont_ptr is noncontiguous and an array but 
not an array section.

In any case, those are restrictions imposed on the user – which the 
compiler may or may not report. (A good one will, I presume. Although, 
one can also regard it as implementation defined/vendor extension as GCC 
will properly handle those – by mapping also the gaps.)

Cheers,

Tobias

> At least the following test case seems to work fine:
>
> integer :: A(10,10), out(12)
> A = reshape([(i, i=0,100)], shape(A))
> !$omp target map(A(3:6,3:5), out)
> !$acc parallel copy(A(3:6,3:5), out)
>   out = reshape(A(3:6,3:5), shape(out))
>   A(3:6,3:5) = 4
> !$acc end parallel
> !$omp end target
> print '(4i3)', out
> print '(10i3)', A
> end
>
> The reason that it works is that the stride is included in the length 
> calculation:
> #pragma omp target num_teams(1) thread_limit(0) map(tofrom:MEM[(c_char 
> *)_6] [len: _5])
>
> Has for the section (with A(3:6,3:5) -> parm(1:4,1:3)):
>               parm.1.dim[0].lbound = 1;
>               parm.1.dim[0].ubound = 4;
>               parm.1.dim[0].stride = 1;
>               parm.1.dim[1].lbound = 1;
>               parm.1.dim[1].ubound = 3;
>               parm.1.dim[1].stride = 10;
> And instead of doing '(4-1+1) * (3-1+1)' = 12 (i.e. multiplying the 
> extends),
> the code does: 'stride * (3-1+1)' = 30.
>
> Cheers,
>
> Tobias
>
> PS: It also works if one puts the stride on the ptr, i.e.:
>
> integer,target :: A(10,10), out(12)
> integer, pointer :: ptr(:,:)
> A = reshape([(i, i=0,100)], shape(A))
> ptr => A(3:6,3:5)
> !$omp target map(ptr, out)
> !$acc parallel copy(ptr, out)
>   out = reshape(ptr, shape(out))
>   ptr = 4
> !$acc end parallel
> !$omp end target
> print '(4i3)', out
> print '(10i3)', A
> end
>
> On 1/4/20 3:25 AM, Julian Brown wrote:
>> This patch tightens up error checking for array references used in
>> OpenACC clauses such that they must now be contiguous. I believe this
>> matches up to the spec (as of 2.6). I've tried to make it so an error
>> only triggers if the compiler is sure that a given array expression must
>> be non-contiguous at compile time, although this errs on the side
>> of potentially allowing the user to shoot themselves in the foot at
>> runtime: at least one existing test in the testsuite appears to expect
>> that behaviour.
>> […]