P2897 R7 aligned_accessor: An mdspan accessor expressing pointer overalignment

[wg21bot]

P2897R0 aligned_accessor: An mdspan accessor expressing pointer overalignment (Mark Hoemmen, Damien Lebrun-Grandie, Nicolas Manual Morales, Christian Trott)

[ben-craig]

2023-10-10 Library Evolution Telecon

P2897R0: aligned_accessor: An mdspan accessor expressing pointer overalignment

2023-10-10 Library Evolution Telecon Minutes

Champion: Mark Hoemmen

Chair: Ben Craig

Minute Taker: Nevin Liber

Start: 2023-10-10

Does this paper have:

• Why does this belong in the standard?
  (From presentation)
  Common vocabulary type for interfaces to declare minimum alignment requirements.
  Extends compilers potential optimizations of assume_aligned.
• Examples? Yes
• Field experience?
  • Implementation experience? Yes
  • Usage experience? Yes
  • Deployment experience? Yes
• Performance considerations? Yes
• Discussion of prior art?
• Changes Library Evolution previously requested? N/A
• Wording? Yes
• Breaking changes? No
• Feature test macro? Yes
• Freestanding? Yes
• Format and/or iostream support? No
• Hash support? No
• Allocator support? No

Summary

POLL: Block aligned_accessor progressing until we have a way of checking alignment requirements during mdspan construction

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
0	1	1	2	2

Attendance: 14

# of Authors: 3

Author Position: WA, SA

Outcome: No consensus to change

WF: concerned about potential for misaligned memory accesses, would like to see something to check alignment

POLL: Mandates (static_assert) vs. Constraints (SFINAE) for converting from low alignment to high alignment. (guidance, feel of the room, non binding if authors discover new info)

Strongly Mandates	Weakly Mandate	Neutral	Weakly Constraints	Strongly Constraints
1	4	1	1	0

Attendance: 14

# of Authors: 3

Author Position: SM, N

Outcome: Preference for Mandates

WM: If there's use cases for SFINAE, then I'd be fine for constraints. Would also like implementer feedback, to ensure there aren't bigger design issues.

Next Steps

Switch to Mandates elements for alignment conversions.

Discuss lack of explicit ctor from default_accessor, as that blocks aligned_mdspan(mdspan).

Authors will explore mdspan safety, in the post-kona timeframe. Change won't be specific to aligned_accessor.

LEWG still needs to review the wording.