Implement C23 memalignment

[Wilco Dijkstra]

Hi Paul,

> I agree that casts are error-prone. However, here the performance of
> memalignment is bad enough that some documentation would be helpful. (In
> Gnulib we never use memalignment, due to this performance hit.)

It's always going to be faster than a division...

> To check whether a pointer @var{P} has alignment @var{A},
> testing @code{(uintptr_t) @var{P} % @var{A}}
> suffices on all but theoretical platforms,
> and because the @code{%} is typically implemented via a simple bitmask

That is incorrect. I just removed very similar code from malloc where it did:

  if ((((unsigned long) (m)) % alignment) != 0)   /* misaligned */

This always uses a division.

> such a test is typically more efficient than
> comparing @code{memalignment (@var{P})} to @var{A}.
> To avoid errors due to overuse of casts,
> the test can be worded as @code{ptr2uint (@var{P}) % @var{A}}
> where @code{ptr2uint} is a trivial inlined function
> whose body contains the cast to @code{uintptr_t}.

I don't think we should ever recommend using '%' since it is too slow.
Writing it as p & (a - 1) would be better since that really does avoid division.
However I still don't see why we would need to document anything at all -
do we document the same for abs() and all other functions that if you switch
off compiler optimization, you might be better off expanding things by hand?

Cheers,
Wilco