C-Array Interoperability of MDSpan
Document #: | PXXXX |
Date: | 2022-02-14 |
Project: | Programming Language C++ LEWG |
Reply-to: |
Christian Trott <crtrott@sandia.gov> Damien Lebrun-Grandie <lebrungrandt@ornl.gov> Mark Hoemmen <mhoemmen@stellarscience.com> K. R. Walker <krwalker@stellarscience.com> Daniel Sunderland <dansunderland@gmail.com> |
Currently one cannot construct an mdspan
from a
multidimensional c-array.
int a[3][2] = {1,2,3,4,5,6};
<int,dextents<2>> b(a,3,2); // fail
mdspan<int,extents<3,2>> c(a); // fail mdspan
It only works for 1D c-arrays, which decay to pointers. However the
deduction guide when using just the 1D c-array with no length argument,
would deduce a rank-0 mdspan
, which may be surprising to
users:
int a[6] = {1,2,3,4,5,6};
(a);
mdspan bstatic_assert(decltype(b)::rank()==0);
We cannot currently fix the multidimensional c-array construction, since it is UB to alias a nested C-Array with a element type pointer - per discussion on the C++ committee reflector in January 2022. However, in practice it works (on the compilers we tested e.g. clang-based and gcc) - and it may be something the committee changes in the future - i.e. make it not-UB. We propotyped this capability, which requires an additional constructor from c-array and a few deduction guides.
What we can fix today is the deduction from 1D c-array, by adding a deduction guide from c-array constraint to rank-1 arrays.
This would enable the following:
{
int a[6] = {1,2,3,4,5,6};
(a);
mdspan bstatic_assert(decltype(b)::rank()==1);
static_assert(decltype(b)::static_extent(0)==6);
}
It is currently possible, and will continue to be allowed to do the following with 1D c-arrays (via the decay to pointer code path):
{
int a[6] = {1,2,3,4,5,6};
(a,3);
mdspan bstatic_assert(decltype(b)::rank()==1);
static_assert(decltype(b)::static_extent(0)==dynamic_extent;
// b.extent(0)==3
}
{
int a[6] = {1,2,3,4,5,6};
<int,extents<3>> b(a);
mdspanstatic_assert(decltype(b)::rank()==1);
static_assert(decltype(b)::static_extent(0)==3;
}
Adding this deduction guide now, means that if we later fix the general issue of constructing an mdspan from a nested c-array (which would require a core language change), we do not break prior behavior.
Based on the reference implementation of P0009 we implemented the changes in a godbolt example: godbolt.
template<class CArray>
(CArray&)
mdspan-> mdspan<see below>;
template<class Pointer>
(Pointer&)
mdspan-> mdspan<see below>;
template<class CArray>
(CArray&)
mdspan-> mdspan<see below>;
Constraints:
is_array_v<CArray>
is true
.rank_v<CArray>==1
is true
.Remarks:
mdspan< std::remove_all_extents_t<CArray>, extents<extent_v<CArray,0>> >
template<class Pointer>
(Pointer&)
mdspan-> mdspan<see below>;
Constraints:
is_array_v<Pointer>
is
false
.
is_pointer_v<Pointer>
is
true
.
Remarks:
The deduced type is:
mdspan< std::remove_pointer_t<Pointer>, extents<> >
Modify paragraph 20 to be:
Constraints:
(is_convertible_v<Integrals, size_type> && ...)
is true
and sizeof...(Integrals) > 0
is
true
.