libstdc++: Handling user-defined interlike object in submdspan.
Tomasz Kaminski
tkaminsk@redhat.com
Tue Sep 30 11:01:19 GMT 2025
On Tue, Sep 30, 2025 at 12:50 PM Luc Grosheintz <luc.grosheintz@gmail.com>
wrote:
>
>
> On 9/30/25 11:44 AM, Tomasz Kaminski wrote:
> > On Tue, Sep 30, 2025 at 11:05 AM Luc Grosheintz <
> luc.grosheintz@gmail.com>
> > wrote:
> >
> >>
> >> On 9/30/25 10:01 AM, Tomasz Kaminski wrote:
> >>>>
> >>>> I canonicalize tuple-like to `std::tuple`s of possibly wrapped
> >>>> index_type; and not to stride<*, *, std::cw<1>>.
> >>>>
> >>>> The advantage is that the preconditions remain the same, because
> >>>> tuples still exist, it's just that we don't encounter as much
> >>>> variability but the kinds of objects remain the same.
> >>>>
> >>>> Generally, and as written in the paper, it does make sense to maybe
> >>>> not canonicalize things like:
> >>>>
> >>>> - std::array<index_type, 2>
> >>>> - std::tuple<std::constant_wrapper<index_type, *>,
> >>>> std::constant_wrapper<index_type, *>>
> >>>>
> >>> Note that pair-like for the submdspan is representing [begin, end)
> range,
> >>> while the strided_slice has [begin, offset, stride)?
> >>>
> >>
> >> Yes, the thought is that
> >> strided_slice{begin, end - begin, std::cw<1>}
> >> is a less direct way of expressing the idea. So users
> >> might opt for one of the other two instead. If we allow
> >> those to be passed through unchanged, they'll not have to
> >> worry about optimizing their code by making sure they
> >> use stride_slice{..., std::cw<1>}, it would just work.
> >>
> > Note, that strided_slice{..., std::cw<1>} may still be passed, so I do
> not
> > see how that removes the number of cases that may be optimized.
> > Canonicalization is strictly reducing the number of cases that user
> defined
> > mappings needs to handle.
> >
>
> True, and not the point I'm trying to make. To me it seems like we've
> put a lot of effort into optimizing small inefficiencies, e.g. removing
> __glibcxx_asserts if they're known to be true at compile time, etc.
>
> Therefore, I was expecting push-back if I simply copy the slices into a
> canonical form. Causing every use of std::array<*, 2> and std::tuple to
> require a copy.
For index_type, passing them by value (copy) should be more efficient than
passing by reference, and the submdspan and standard layouts take them by
value,
so we copy std::array or std::tuple anway (see
https://eel.is/c++draft/mdspan.sub.map).
I do not see much difference between copying two ints by passing
std::array<int, 2>
by value or constructing prvalue strided_slice<int, int,
constant_wrapper<1>>.
std::tuple has a user-provided assignment operator, so it is not trivially
copyable, thus
will not be passed by registers.
Note, that for the implementation, I would not use
submdspan_canonicalize_indicies,
that creates a tuple. And instead do something like:
submdspan(SliceSpecifiers... slices)
{
if constexpr (!is_cannonical_slice<SliceSpecifiers> || ....)
return submdspan(canonicalize(slices)....);
else
{
// Now all slices are cannonica.
}
}
> I've recently seen the GCC optimizer perform small
> miracles, so canonicalize first and then add exceptions as needed makes
> sense to me.
>
> To make the above more meaningful: I don't believe that users would
> express std::array{being, end} as:
>
> std::strided_slice{begin, end - begin, std::cw<1>}.
>
> Hence, the assumption that exceptions for std::array and std::tuple
> make sense (for the sake of saving copies).
>
Copies are still done, as we define the friend function as passing by value.
Or do you mean something else here?
> >
> >>
> >>>
> >>>>
> >>>> because those are very natural ways of specifying that kind of slice.
> >>>> Therefore, even if there's a cost to canonicalization, every kind of
> >>>> slice has a convenient canonical form (that users can provide to skip
> >>>> canonicalization).
> >>>>
> >>> I do not think there should be differences between canonical forms of
> >>> strided_slice and pair_like, they represent the same concept. And
> >>> This should reduce the number of various concepts that we need to
> handle.
> >>> Note that the __first, __last functions are now completely gone from
> the
> >>> wording.
> >>
> >> I agree that multiple canonical forms is odd. On the other hand
> >> it's mentioned/suggested in the paper that standardized mapping
> >> may do less canonicalization:
> >>
> >>
> >>
> https://isocpp.org/files/papers/D3663R3.html#submdspan-can-skip-canonicalization-for-standard-layout-mappings
> >
> > Standard mapping are required to strided_slice{..., std::cw<1>} anyway,
> > and
> > I would prefer to avoid having two code paths, one for standard mapping
> and
> > other
> > for user-defined mappings, if that is not proven strictly necessary.
> >
> >
> >>
> >>>
> >>>
> >>>>
> >>>> Is it okay to send this variation as a v1? Naturally, you can then
> >>>> say: nope, do the full P3663 please.
> >>>>
> >>> You can send that of course, but after rewriting the whole wording
> >> section,
> >>> I think the previous
> >>> specification was a bit of a dead end, so I think I wil answer that way
> >>> anyway.
> >>>
> >>>
> >>>>>
> >>>>>
> >>>>>>
> >>>>>>>>
> >>>>>>>> The idea I'm pursuing is to follow the pattern of casting slices
> to
> >>>>>>>> "standard slices" immediately at the user-facing API layer, e.g.
> >>>>>>>>
> >>>>>>>> template<typename _Slices...>
> >>>>>>>> constexpr auto
> >>>>>>>> submdspan_extents(_Slices... __slices)
> >>>>>>>> {
> >>>>>>>> return
> >>>> __submdspan_extents(__slice_cast(std::move(__slices))...);
> >>>>>>>> }
> >>>>>>>>
> >>>>>>
> >>>>>>
> >>>>>
> >>>>
> >>>>
> >>>
> >>
> >>
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://gcc.gnu.org/pipermail/libstdc++/attachments/20250930/d6233ef4/attachment-0001.htm>
More information about the Libstdc++
mailing list