not_null cannot (and should not) be used with span

[hpesoj]

F.18 suggests that not_null can be used with span and string_span, but my tests would suggest otherwise. This behaviour seems correct, as arrays and strings cannot be null (null std::array or std::string anyone?) Indeed, span's operator bool() tests for "empty" status, and trying to compare a span with nullptr results in a compiler error.

[hsutter]

Discussed on 12-8. We are still are inclined to enable not_null be usable with span but will discuss further.

[AndrewPardoe]

Editors discussed 8 Dec 2015.

[hpesoj]

Thanks for your consideration.

As I just mentioned in microsoft/GSL#202, if you insist on legitimising the concept of a "null span", you could introduce a special nullspan constant to use instead of nullptr. This would:

• allow better overload resolution between pointers and spans;
• make it clear that references/pointers and arrays are conceptually different things; and
• still make grammatical sense with the not_null annotation.

I consider the concept of a "null array" to be conceptually unsound, but this seems like it would be a reasonable compromise.

[AndrewPardoe]

Editors discussed 14 Dec 2015

[AndrewPardoe]

Herb, please write this up. Per our discussion, suggest that we don't allow not_null for smart pointers, including span<T>. Even if we had not_null<unique_ptr<T>>,it wouldn't be movable. Checking span<T> against null is problematic anyway because you should be checking against empty.