[hardware.interference] Bad examples of using alignas

[Halalaluyafail3]

Example 1 is the following:

struct keep_apart {
  alignas(hardware_destructive_interference_size) atomic<int> cat;
  alignas(hardware_destructive_interference_size) atomic<int> dog;
};

There is nothing specifying that these must be valid alignments, or that they must even be nonnegative integer powers of two. Would it make sense to use something like bit_ceil here, plus a check to handle the case where that result is not representable? That would fix the latter issue, not sure if it is possible to check if something is a valid alignment.

Example 2 is the following:

struct together {
  atomic<int> dog;
  int puppy;
};
struct kennel {
  // Other data members...
  alignas(sizeof(together)) together pack;
  // Other data members...
};
static_assert(sizeof(together) <= hardware_constructive_interference_size);

Why is alignas(sizeof(together)) being done? The two issues mentioned above are also present here.

[jensmaurer]

This feels like LWG issue material, not editorial.

[Halalaluyafail3]

Is it necessary for LWG to change the examples? They are not normative so I would expect that they can be changed editorially.

[eisenwave]

It's pretty clear that P0154R1 wants these constants to be usable in alignas, which necessitates powers of two. If the wording doesn't make that clear, that should be done as an LWG issue.

I guess alignas(sizeof(together)) is making sure that a together object goes into a cache line instead of being split up. An object whose alignment is equal to its size cannot be split if the access is aligned. To be fair, the example makes some assumptions about struct padding and int size; it's not 100% guaranteed that the size is a power of two, especially not if int is 3 bytes large or something else weird. Maybe we should document those assumptions.

[jwakely]

Right, the point of those constants is to specify alignment. If your C++ implementation defines them with values that aren't valid alignments, use a better implementation. Being required to use bit_ceil would be absurd.