How my papers did at Issaquah

The ISO C++ Committee met in Issaquah, Washington, last week. I currently have three papers in the pipeline:

How did they fare?

Static storage for braced initializers

P2752 “Static storage for braced initializers” was seen by the EWG Incubator on Friday night, and passed on to EWG uneventfully, as expected. Since then, its proposed wording has been slightly wordsmithed (thanks, Jens Maurer!) so that hopefully it’ll be ready for C++26 at the next meeting. If it is accepted, you’ll be able to write (Godbolt)

int f(std::initializer_list<int>);
int test() {
    return f({1,2,3,4,5});

and have the compiler (1) use no stack space for that function, and (2) make the call to f a tail-call. Right now, due to an oversight in the C++11 rules for initializer_list, this doesn’t work.

As you can see on Godbolt, I have a prototype implementation of P2752 under the Clang command-line flag -fstatic-init-lists.

You might further ask, “I thought the vocabulary type for this kind of thing, since C++20, was span. Why does the following snippet not even compile yet?” (Godbolt.)

int g(std::span<const int>);
int test() {
    return g({1,2,3,4,5});

That problem would be fixed by the adoption of Federico Kircheis’ P2447R2, which you may remember from std::span should have a converting constructor from initializer_list (2021-10-03). I don’t believe there was any movement on P2447 at Issaquah; in fact I’ve heard that it is dead (perhaps because it presented as a change to C++26 instead of as a defect against C++20); but I would love to see it revived.

GCC -fmerge-all-constants

UPDATE, 2023-02-25: Kevin Zheng informs me that GCC trunk already has an even more powerful flag, -fmerge-all-constants. That flag does things the Standard will never in a million years condone, such as merging the constant arrays in

void h(const int*, const int*);
void test() {
    const int a1[] = {1,2,3,4,5};
    const int a2[] = {1,2,3,4,5};
    h(a1, a2);

GCC’s -fmerge-all-constants produces optimal codegen for initializer_list — even more optimal than my quick-and-dirty prototype of -fstatic-init-lists! — but at the cost of non-conforming codegen for the above.

Clang also has -fmerge-all-constants, but for historical reasons it is much less powerful. Historically, -fmerge-all-constants was Clang’s default behavior. So it was watered down to avoid many of the obvious dangers; for example, it would never merge a1 and a2 above. Still, people had been complaining about the non-conforming default since at least 2014; and they finally turned it off entirely in Clang 6.0.1 (July 2018). Passing -fmerge-all-constants will still re-enable the non-conforming behavior — but only in the watered-down, almost-conforming state that it was in when it was turned off. Clang has nothing like GCC’s full-strength behavior.

Improving std::hive

P2596 “Improve std::hive::reshape proposes simplifications to P0447R21 std::hive that would give vendors more freedom to optimize, shrink its memory footprint, and make the hive::splice member function noexcept. P2596 was seen by LEWG on Tuesday, with the following straw-poll results:

Do we ever want to see P0447 std::hive again? 8 4 5 2 3
Do we want to see P0447 std::hive again with P2596’s changes applied? 2 3 8 4 3

Committee chairs are fond of describing straw-poll results as “tea leaves,” and (in order to self-fulfill that prophecy) seem to delight in crafting polls that convey as little concrete information as possible. One way to read this pair of polls is that the combined P0447+P2596 had far less support than the status quo: 12–5–5 in favor of P0447, 5–8–7 in favor of P0447+P2596 specifically. Another way to read the polls is that when you subtract out the five people in the room who never want to see P0447 again in any form whatsoever, the room is 5–8–2 in favor of applying P2596’s changes.

I’ve made a “not-really-production-quality” implementation of P0447 std::hive available on Godbolt. The P2596 API is available under the macro -D_LIBCPP_P2596.

Trivial relocatability

P1144 “std::is_trivially_relocatable has a new name! I’ve decided that its former name, “Object relocation in terms of move plus destroy,” was a little too much of a mouthful; and everyone knows the paper as “Trivially Relocatable” anyway.

There was big news on the trivial-relocation front at Issaquah: Bloomberg has entered the fray with a paper coauthored by Mungo Gill and Alisdair Meredith, P2786R0 “Trivial relocatability options.” This paper wasn’t finished in time for the pre-Issaquah mailing and was continually revised all week leading up to Friday night’s discussion; it obeys Pascal’s dictum, weighing in at 30 pages. By comparison, P1144R6 is 17 pages; my draft of P1144R7 is down to 15 pages so far, and I’m trying hard to shorten it.

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.”

EWGI straw-polled only this third bullet point, and as a three-way poll rather than the usual five-way; the results were a dismal 7–5–6 in favor of optimizing.

But there is one major point where P2786R0’s disagreement with P1144R6 was wise and good!

  • Whether the std::uninitialized_relocate algorithm should do the Slower But Right Thing when the source and destination ranges overlap, as std::memmove does; or whether it should do the Fast But Wrong Thing, as std::memcpy and std::copy do. (Godbolt.) P1144 says “fast but wrong,” P2786 says “slower but right.”

The trouble is, you can really only do the Right Thing on overlap if you can detect overlap, and that’s possible only if you are given contiguous iterators. Qt defines two different algorithms here: q_uninitialized_relocate_n whose trivial path can use memcpy (EDIT: now it does), and q_relocate_overlap_n whose trivial path must use memmove. Should the standard library do the same?

In the next mailing, you can expect to see some changes in P1144R7 — beyond just shortening the darn thing! — and probably a new joint paper coauthored by Alisdair Meredith and myself that expands further on the specific design-decision bullet points above (and maybe some others, too).

Expect further blog posts in the next few weeks about trivial relocatability. If you’re champing at the bit, see my existing posts tagged #relocatability, and Dana Jansens’ blog post “Trivially Relocatable Types in C++/Subspace” (January 2023). (EDIT: Here are the promised posts: I, II, III.)

Posted 2023-02-17