This page is a snapshot from the LWG issues list, see the Library Active Issues List for more information and the meaning of New status.

3597. Unsigned integer types don't model advanceable

Section: [range.iota.view] Status: New Submitter: Jiang An Opened: 2021-09-23 Last modified: 2021-10-17

Priority: 3

View other active issues in [range.iota.view].

View all other issues in [range.iota.view].

View all issues with New status.


Unsigned integer types satisfy advanceable, but don't model it, since

every two values of an unsigned integer type are reachable from each other, and modular arithmetic is performed on unsigned integer types,

which makes the last three bullets of the semantic requirements of advanceable ( [range.iota.view]/5) can't be satisfied, and some (if not all) uses of iota_views of unsigned integer types ill-formed, no diagnostic required.

Some operations that are likely to expect the semantic requirements of advanceable behave incorrectly for unsigned integer types. E.g. according to [range.iota.view]/6 and [range.iota.view]/16, std::ranges::iota_view<std::uint8_t, std::uint8_t>(std::uint8_t(1)).size() is well-defined IMO, because

Bound() is std::uint8_t(0), which is reachable from std::uint8_t(1), and not modeling advanceable shouldn't affect the validity, as both W and Bound are integer types.

However, it returns unsigned(-1) on common implementations (where sizeof(int) > sizeof(std::uint8_t)), which is wrong.

Perhaps the semantic requirements of advanceable should be adjusted, and a refined definition of reachability in 24.6.4 [range.iota] is needed to avoid reaching a from b when a > b (the iterator type is also affected).

[2021-10-14; Reflector poll]

Set priority to 3 after reflector poll.

[Tim Song commented:]

The advanceable part of the issue is NAD. This is no different from NaN and totally_ordered, see [structure.requirements]/8.

The part about iota_view<uint8_t, uint8_t>(1) is simply this: when we added "When W and Bound model ..." to [range.iota.view]/8, we forgot to add its equivalent to the single-argument constructor. We should do that.

Proposed resolution: