[Bug libstdc++/78420] [5/6/7 Regression] std::less<T*> is not a total order with -O2 enabled

[jason at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78420

--- Comment #15 from Jason Merrill <jason at gcc dot gnu.org> ---
(In reply to Ville Voutilainen from comment #14)
> Not in general, no, it doesn't have to always give a compile-time answer.
> But I believe the library intent is that when it compares compile-time
> constant pointers, it should give that answer at compile-time, and it should
> be a total order.

Right.

> What seems to be suggested by that comment is that since
> it's possible to compare compile-time constant pointers and run-time values
> with this function, and it can't know (definition-wise) whether the incoming
> arguments are constants or not, we could just as well drop the constexpr
> completely from the pointer specialization of std::less, as that will then
> open the door to simply reinterpret_cast in it. None of that has been
> confirmed by LWG, though.

I think this is an example of the prohibition of reinterpret_cast being too
broad.

On the other hand, compile-time constant pointers aren't of much use.

On the third hand, using std::less within an array defined in a constexpr
function seems entirely reasonable and likely to come up regularly.

I wonder about instead relaxing the total order requirement somewhat, to
something like saying that repeated comparisons must be consistent.