LEWG 2026/02/10: PL-007, P3982R0

[mhoemmen]

PL-007 23.7.3.7 [mdspan.sub] Define the extent member of the strided_slice

NB comment: https://github.com/cplusplus/nbballot/issues/816

Paper: https://isocpp.org/files/papers/P3982R0.html

Make strided_slice:::extent define the output extent, not the input span

2 ranges of indices in submdspan

1. input span: range of indices into input, that can be accessed by output
2. output extent: size of the range of valid indices for output

2 possible choices of meaning for strided_slice's extent

1. input span $[$ offset, offset + extent $)$, thus output extent is 1 + (extent - 1) / stride
   • STATUS QUO
2. output extent, thus input span is $[$ offset, 1 + offset + (extent - 1)*stride $)$
   • P3982

Example: status quo vs. P3982R0

Example: strided_slice{.offset = 1, .extent = 10, .stride = 3}.

• Status quo
  • View input indices 1, 4, 7, and 10
  • Input span: $[1, 1 + 10) = [1, 11)$
  • Output extent: offset + (extent - 1) / stride = 1 + (10 - 1) / 3 = 4
    • NOT same as input .extent 10

• P3982R0
  • View input indices 1, 4, 7, 10, 13, 16, 19, 22, 25, 28
  • Input span: $[$ offset, offset + 1 + (extent - 1)*stride $)$ = $[1, 29)$
  • Output extent: 10
    • Same as input .extent

Status quo aimed for minimal change from first : last : stride

• Original submdspan paper P2630 lacked P3663's notion of "canonical slice types" vs. all slice types
• Thus, P2630 authors defined slice types based on familiarity
• Most familiar syntax: Fortran's and Python's first : last : stride
• But first : last : stride can't represent (dynamic offset, static extent, static stride)
  • Key case for performance, e.g., unrolled loop
• Status quo in P2630 made the minimal-distance-from-familiarity change to include this case:
  • Replace last with extent = last - first
  • Rename first to offset

Why change status quo?

NB comment points out 2 reasons:

1. Permit zero strides in the future (for "broadcasting" layouts)
2. Enable representing (static extent, dynamic stride)

P3982R0 points out 2 more:

3. Avoid integer divisions
   • Performance
   • Remove undefined behavior (division by zero stride)
4. Less confusing (input "extent" is not output extent; I have to look this up every time)

Another reason: strided_slice is a canonical slice type

• Canonical slice types define interface between submdspan and a layout mapping's submdspan_mapping
• submdspan function body looks like this:

auto [...canonical_slices] =
    submdspan_canonicalize_slices(src.extents(), user_slices...);
auto sub_map_result =
    submdspan_mapping(src.mapping(), canonical_slices...);
return mdspan(
    src.accessor().offset(src.data_handle(), sub_map_result.offset),
    sub_map_result.mapping,
    typename AccessorPolicy::offset_policy(src.accessor()));

• Choose canonical types to express the biggest possible set of slices
• We can always add new user-facing slices to improve familiarity (see range_slice below)

Why do we need to change it for C++26?

Changing the meaning of strided_slice::extent later would silently break the meaning of submdspan.

(It would be an ABI break. You don't like those, right?)

Also for C++26: Rename strided_slice to extent_slice

• Paper proposes a new non-canonical slice type range_slice to cover common use case first : last : stride (Python, Fortran, etc.)

• Whether that arrives in C++26 or C++29, that would mean there are multiple *_slice types with a stride

• strided_slice is a canonical slice type; changing name after C++26 would break users' customizations of submdspan_mapping

• New name: extent_slice, because it has the output extent, not the input span

Additional new features

Paper proposes 2 new features. They could be added for C++26 or in a later Standard.

1. range_slice{.first = first, .last = last, .stride = stride}

   • Models slicing in other programming languages
   • Calls for renaming strided_slice TO extent_slice NOW, even if we add range_slice later
     • range_slice also has a "stride"
     • So "strided_slice" wouldn't be the only "strided" slice

2. Interpret any type for which structured binding auto [first, last, stride] = s; is valid as range_slice

   • Only makes sense if we add range_slice as in (1)

Why add range_slice in C++26?

• Popular programming languages have first : last : stride slices, so users will demand them

  • Fortran: array(first:last), array(first:last:step)
  • Python: array[first:last], array[first:last:step]
  • Matlab: array(first:last), (annoyingly) array(first:step:last)

• Designated initializers alleviate order concerns

Why add auto [first, last, stride] = s;?

• Confusion about order (Python vs. Matlab) led the submdspan (P2630) authors not to include this feature, but
  • status quo already interprets auto [first, last] = s; as stride = cw<1zu>
  • so auto [first, last, stride] = s; is less ambiguous, analogous to a default function parameter stride = cw<1zu>
• Trivial to implement: we already handle auto [...ts] = t; with sizeof...(ts) == 2
• We consider this optional for C++26

Suggested polls

1. Accept rename [to extent_slice] and changes to strided_slice class template from P3982R0 to C++26.

2. Accept the introduction of range_slice class template from P3982R0 to C++26.

3. Accept any type decomposable into three elements as submdspan slice type as proposed in P3982R0 to C++29.

4. Accept any type decomposable into three elements as submdspan slice type as proposed in P3982R0 to C++26.

[zhanglin2603]

Hey, I went over #448 and can take this on.

• My read: this is primarily a core implementation task tied to "LEWG 2026/02/10: PL-007, P3982R0".
• I will focus on these issue-specific points first: lewg, pl-007, p3982r0.
• First I will reproduce and confirm expected behavior for the core implementation path.
• Then I will implement the fix with targeted tests and keep scope tight (good confidence).
• I will share a working PR update fast (24h for first patch, 48h for final polish) with verification notes.
  If this approach works for you, I can start now.