CWG3004 [expr.const] Pointer arithmetic on an array whose bound is not yet known

[frederick-vs-ja]

Full name of submitter (unless configured in github; will be published with the issue): Barry Revzin

Reference (section label): [expr.const]

Link to reflector thread (if any):

Issue description:

From P2280R4:

Perhaps the most fun example is this one:

extern const int arr[];
constexpr const int *p = arr + N;
constexpr int arr[2] = {0, 1};
constexpr int k = *p;

Which every compiler currently provides different results (in order of most reasonable to least reasonable):

1. Clang says arr+N is non-constant if N != 0, and accepts with N == 0.
2. GCC says arr+N is always constant (even though it sometimes has UB), but rejects reading *p if arr+N is out of bounds.
3. ICC says arr+N is always constant (even though it sometimes has UB), but always rejects reading *p even if arr+N is in-bounds.
4. MSVC says you can’t declare arr as non-constexpr and define it constexpr, even though there is no such rule

This, to me, seems like there should be an added rule in [expr.const] that rejects addition and subtraction to an array of unknown bound unless that value is 0. This case seems unrelated enough to the rest of the paper that I think it should just be a Core issue.

It seems that no CWG issue is created for this yet.

Suggested resolution:

[frederick-vs-ja]

CC @brevzin

[t3nsor]

There's another gap as well in the wording for arrays of unknown bound. We say in [basic.types.general]

The declared type of an array object can be an array of unknown bound and therefore be incomplete at one point in a translation unit and complete later on; the array types at those two points (“array of unknown bound of T” and “array of N T”) are different types.

which heavily hints, but does not outright normatively specify, that at a point from which declarations are reachable that give a variable both types "array of unknown bound of T" and "array of N T", the latter is considered to be the type.

[jensmaurer]

CWG3004