[basic.align] Clarify alignment requirements of non-static data members

[frederick-vs-ja]

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

Reference (section label): [basic.align]

Link to reflector thread (if any):

Issue description:

(from cplusplus/draft#8787)

Currently, the standard wording doesn't directly specify the alignment requirement of a non-static data member.

Consider the following example.

struct S1 { int n; };
struct S2 { alignas(int) int n; };

It's unclear whether the alignment requirement of the n member of S1 is same as that of its type. If not so, then alignas(int) in S2 strengthens alignment requirements, and S1 and S2 are not layout-incompatible.

Also, consider the following trickier example.

struct W1 { char c; }; // assume that alignof(W1) == 1
struct alignas(2) W2 { char c; };

struct alignas(4) S3 { W1 w; };
struct alignas(4) S4 { W2 w; };

In this example, for each S3 or S4 object o, its o.w member subobject is indirectly required to be located at an address required by alignas(4), because o.w must have the same address as o. Can we infer that such a w member has a stricter alignment requirement? I don't think so, but the current wording doesn't clearly rejects such reading.

Suggested resolution:

Modify [basic.align] as indicated:

1. [...] An object type imposes an alignment requirement on every object or non-static data member of that type; stricter alignment can be requested using the alignment specifier ([dcl.align]). [...]
   [Note ?: It is possible for a member subobject to have stricter restrictions on its locatable address due to the alignment requirement of its enclosing class object. However, such restrictions do not affect the alignment requirement of the corresponding non-static data member. — end note]

[languagelawyer]

1. An object type imposes an alignment requirement on every object or non-static data member of that type

Member subobjects are objects. This change changes nothing except that it mixes 2 «different level» concepts: objects and NSDMs. (The change could make sense if the paragraph was saying «on every variable of that type», but the more correct way woulda be just replacing «variable» with «object»)

[frederick-vs-ja]

Member subobjects are objects. This change changes nothing except that it mixes 2 «different level» concepts: objects and NSDMs. (The change could make sense if the paragraph was saying «on every variable of that type», but the more correct way woulda be just replacing «variable» with «object»)

Member subobjects are objects, but IIUC a NSDM is not itself a member subobject. Currently, [class.mem.general] p7 says that a non-static data member corresponds to member subobjects, and it should be clear that they're not the same thing.

Also, [class.mem.general] p27 talks about NSDMs but not subobjects, so I think we may need to directly specify alignment requirements of NSDMs.

[languagelawyer]

Member subobjects are objects

So, they are already covered by «An object type imposes an alignment requirement on every object of that type» then?

IIUC a NSDM is not itself a member subobject

Correct.

[class.mem.general] p7 says that a non-static data member corresponds to member subobjects

It could be more clear about subobject type, maybe it is specified somewhere else, but anyway.

[class.mem.general] p27 talks about NSDMs but not subobjects, so I think we may need to directly specify alignment requirements of NSDMs

It feels that [class.mem.general]/27 needs a fix, not the description of alignment requirement. E.g. by saying «corresponding entities have the same compatible alignment requirements specifiers» (with proper definition of compatible alignment specifiers)

[jensmaurer]

We're often a bit sloppy observing the subobject / non-static data member boundary, but we do really want to talk about the resulting (effective) alignment requirements, not about any alignment specifiers that happen to be hanging around.

So, "int x" and "alignas(int) int y" have the same alignment requirements, and are indistinguishable from that angle.