[diff.cpp20.thread]

# Annex C (informative) Compatibility [[diff]](./#diff)

## C.2 C++ and ISO C++ 2020 [[diff.cpp20]](diff.cpp20#thread)

### C.2.13 [[thread]](thread "32 Concurrency support library"): concurrency support library [diff.cpp20.thread]

[1](#1)

[#](http://github.com/Eelis/draft/tree/9adde4bc1c62ec234483e63ea3b70a59724c745a/source/compatibility.tex#L797)

**Affected subclause:** [[thread.barrier]](thread.barrier)  


**Change:** In this revision of C++,
it is implementation-defined whether a barrier's phase completion step runs
if no thread calls wait[.](#1.sentence-1)

Previously the phase completion step was guaranteed to run on the last thread that calls arrive or arrive_and_drop during the phase[.](#1.sentence-2)

In this revision of C++,
it can run on any of the threads that arrived or waited at the barrier
during the phase[.](#1.sentence-3)

  


**Rationale:** Correct contradictory wording and
improve implementation flexibility for performance[.](#1.sentence-4)

  


**Effect on original feature:** Valid C++ 2020 code using a barrier might have
different semantics in this revision of C++
if it depends on a completion function's side effects occurring exactly once,
on a specific thread running the phase completion step, or
on a completion function's side effects occurring
without wait having been called[.](#1.sentence-5)

[*Example [1](#example-1)*: auto b0 = std::barrier(1);
b0.arrive();
b0.arrive(); // implementation-defined; previously well-definedint data = 0;auto b1 = std::barrier(1, [&] { data++; });
b1.arrive();
assert(data == 1); // implementation-defined; previously well-defined b1.arrive(); // implementation-defined; previously well-defined — *end example*]