On Fri, Mar 15, 2024 at 11:51 AM Ville Voutilainen <ville.voutilainen@gmail.com> wrote:
On Fri, 15 Mar 2024 at 15:12, Arthur O'Dwyer <arthur.j.odwyer@gmail.com> wrote:

>> >> What is a user of P1144 expected to do about
>> >> "pmr and allocator aware types will behave in
>> >> surprising ways when used in containers and
>> >> algorithms" ?
>> >
>> > The proof is in the pudding! Try to find me a "surprising" behavior of allocator-aware types. The complete implementation of P1144, one compiler and two standard libraries, is available for perusal.
>> > - https://godbolt.org/z/eM8saKe6b
>> I take that as a suggestion to stop spending time on P1144 and to
>> focus all work on P2786 instead.
>> I do not have time for reverse-engineering riddles.
> And yet you have time to troll the SG14 mailing list.

You aren't qualified to call other people's questions trolling.

> Look, you made the claim. Please provide your evidence.

What claim? I was quoting YOUR PAPER. A statement made in P2814R1.

That's a joint paper with the Bloomberg authors. You'd have to ask them.
I've already quoted you my own (less-than-15-page) summary of the key differences between P1144 and P2786. Here it is again.

The major places where P2786R0 disagreed with P1144 are:

  • Whether std::swap can be implemented as “relocate A into a temporary; relocate B into A; relocate the temporary into B,” or whether it must be implemented in terms of assignment — and what “assignment” means for trivial types anyway. Vendors already optimize copy_n in a way detectable by the user; the question is if we should permit vendors to optimize swap in the same way. P1144 says “optimize,” P2786 says “don’t.”

  • Whether vector::insert can use relocation to “make a window” that is then filled with the new data (as in Folly’s fbvector), or whether it must be implemented in terms of assignment. P1144 says “optimize,” P2786 says “don’t.” See “Should assignment affect is_trivially_relocatable?” (2024-01-02).

  • Whether a type containing a data member of type boost::interprocess::offset_ptr<T> may be explicitly warranted as [[trivially_relocatable]] (Godbolt). P1144 says “optimize,” P2786 says “don’t.”