P2599 mdspan::size_type should be index_type

[brycelelbach]

P2599R0 mdspan::size_type should be index_type (Nevin Liber)

[brycelelbach]

2022-06-07 Library Evolution Telecon

P2599R0: mdspan::size_type should be index_type

2022-06-07 Library Evolution Telecon Minutes

Chair: Bryce Adelstein Lelbach

Minute Taker: Ben Craig

Champion: Nevin Liber

Start: 2022-06-07 12:03 Eastern

POLL: Send P2599R0 (mdspan::size_type should be index_type) to Library for C++23 classified as an improvement (B2), to be confirmed with a Library Evolution electronic poll.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
2	7	2	2	1

Attendance: 21

# of Authors: 1

Author Position: SF

Outcome: No consensus.

SA: It's already a conscious choice by the user to use a signed type. So I don't think it will be surprising. The consistency of having it be called size_type is more important.

POLL: mdspan, extents, and layouts should have both an index_type (which is whatever the user provides for the first template parameter to extents) and a size_type (which is make_unsigned_t<index_type>).

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
3	9	1	1	0

Attendance: 19

# of Authors: 1

Author Position: SF

Outcome: Strong consensus in favor, and stronger consensus than the paper as written.

WA: It's additional complexity.

End: 12:13

Summary

This paper proposes that we should rename mdspan's size_type to index_type, now that the index type is parameterized and may be either unsigned or signed.
Some felt that it might be surprising to users for mdspan to not use the familiar size_type name.
Others felt that it would be surprising for size_type to be an alias for a signed type, which is the status quo.
We decided that mdspan should have an index_type, which is the index type provided by the user, and also a size_type, which is make_unsigned_t<index_type>.

Next Steps

Produce and publish P2559R1 (mdspan::size_type should be index_type), making index_type an alias for the user-provided index type and size_type an alias for make_unsigned_t<index_type>.

[brycelelbach]

2022-06-21 Library Evolution Telecon

P2599R1: mdspan::size_type should be index_type

2022-06-21 Library Evolution Telecon Minutes

Chair: Bryce Adelstein Lelbach

Minute Taker: Ben Craig

Champion: Nevin Liber

Start: 2022-06-21 11:15 Eastern

Does this paper have:

• Changes Library Evolution previously requested?
  • Yes

Give the paper a clearer name now that the design intent has changed to prevent any confusion.

What should mdspan::size return?

• P0009R16: size_type.
• P0009R17 (after we parameterized index type): size_t.
• P2599R1: index_type.

mdspan::size should probably return size_type.

What about layout mappings and mdpan's required_span_size?
This should remain index_type for performance.

What about layout mappings operator()?
The result of that is fed into accessors, which take a size_t.

Why do accessors take size_t instead of index_type?

Why doesn't layout mapping use operator[] instead of operator()?
Because it's a function object.

POLL: Modify P2599R1 (mdspan::size_type should be index_type) such that mdspan::sizes return type is size_type, and send the modified paper to Library for C++23 classified as B2 - Improvement, to be confirmed with a Library Evolution electronic poll.

Strongly Favor	Weakly Favor	Neutral	Weakly Against	Strongly Against
5	8	0	1	0

Attendance: 29

# of Authors: 1

Author Position: SF

Outcome: Strong consensus in favor.

WA: This is a late change.

End: 11:41

Summary

This paper proposes that mdspan should have an index_type, which is the index type provided by the user, and also a size_type, which is make_unsigned_t<index_type>.

We discussed what mdspan::size should return.
Before mdspan was parameterized on index type, it returned size_type.
After that change, it returned size_t.

The paper proposed that mdspan::size should be changed to return index_type.
We discussed this and believe that the new size_type is a more suitable return type.
We also discussed whether required_span_size on layouts should return index_type or
size_type, and decided that it should return index_type for performance.
We also considered what type accessors should take and return.
Currently, accessors are not parameterized on the index type, and expect and return size_t.

We forwarded the paper, with the modification of mdspan::size returning size_type, to Library for C++23, pending confirmation with an electronic poll.

Next Steps

• Produce and publish P2599R2 (mdspan::size_type should be index_type), improving the paper title and changing mdspan::size's return type to size_type.
• Take an electronic poll to confirm sending P2599R2 (mdspan::size_type should be index_type) to Library for C++23 classified as B2 - Improvement.

[wg21bot]

P2599R1 mdspan::size_type should be index_type (Nevin Liber)