Merge 2023-02 LWG Motion 19

P2588R3 barrier's phase completion guarantees

Author: tkoeppe

source/compatibility.tex

@@ -322,6 +322,40 @@
 };
 \end{codeblock}
 
+\rSec2[diff.cpp20.thread]{\ref{thread}: concurrency support library}
+
+\diffref{thread.barrier}
+\change
+In this revision of \Cpp{},
+it is implementation-defined whether a barrier's phase completion step runs
+if no thread calls \tcode{wait}.
+Previously the phase completion step was guaranteed to run on the last thread that calls \tcode{arrive} or \tcode{arrive_and_drop} during the phase.
+In this revision of \Cpp{},
+it can run on any of the threads that arrived or waited at the barrier
+during the phase.
+\rationale
+Correct contradictory wording and
+improve implementation flexibility for performance.
+\effect
+Valid \CppXX{} code using a barrier might have
+different semantics in this revision of \Cpp{}
+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 \tcode{wait} having been called.
+For example:
+\begin{codeblock}
+auto b0 = std::barrier(1);
+b0.arrive();
+b0.arrive();            // implementation-defined; previously well-defined
+
+int 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{codeblock}
+
 \rSec1[diff.cpp17]{\Cpp{} and ISO \CppXVII{}}
 
 \rSec2[diff.cpp17.general]{General}

source/support.tex

@@ -573,7 +573,7 @@
 #define @\defnlibxname{cpp_lib_atomic_shared_ptr}@                 201711L // also in \libheader{memory}
 #define @\defnlibxname{cpp_lib_atomic_value_initialization}@       201911L // also in \libheader{atomic}, \libheader{memory}
 #define @\defnlibxname{cpp_lib_atomic_wait}@                       201907L // also in \libheader{atomic}
-#define @\defnlibxname{cpp_lib_barrier}@                           201907L // also in \libheader{barrier}
+#define @\defnlibxname{cpp_lib_barrier}@                           202302L // also in \libheader{barrier}
 #define @\defnlibxname{cpp_lib_bind_back}@                         202202L // also in \libheader{functional}
 #define @\defnlibxname{cpp_lib_bind_front}@                        201907L // also in \libheader{functional}
 #define @\defnlibxname{cpp_lib_bit_cast}@                          201806L // also in \libheader{bit}

source/threads.tex

@@ -9643,15 +9643,11 @@
   The expected count is decremented
   by each call to \tcode{arrive} or \tcode{arrive_and_drop}.
 \item
-  When the expected count reaches zero, the phase completion step is run.
-  For the specialization
-  with the default value of the \tcode{CompletionFunction} template parameter,
-  the completion step is run
-  as part of the call to \tcode{arrive} or \tcode{arrive_and_drop}
-  that caused the expected count to reach zero.
-  For other specializations,
-  the completion step is run on one of the threads
-  that arrived at the barrier during the phase.
+  Exactly once after the expected count reaches zero, a thread
+  executes the completion step during its call
+  to \tcode{arrive}, \tcode{arrive_and_drop}, or \tcode{wait},
+  except that it is \impldef{barrier phrase completion without \tcode{wait}}
+  whether the step executes if no thread calls \tcode{wait}.
 \item
   When the completion step finishes,
   the expected count is reset