Export of internal Abseil changes

--
7b6a68aa92dcc7247236d1a1813914e035383bf8 by Abseil Team <absl-team@google.com>:

Use atomic exchange to mark completion in absl::once_flag

This prevents a potential for a missed wakeup if one thread marks
itself as a waiter while another thread is completing the invocation.

PiperOrigin-RevId: 344946791

--
ddff21d1dde08d1368d8be5fca81b154e78be2fc by Abseil Team <absl-team@google.com>:

Add missing string_view include. This is currently used transitively through the cord header.

PiperOrigin-RevId: 344845266
GitOrigin-RevId: 7b6a68aa92dcc7247236d1a1813914e035383bf8
Change-Id: Ia24e98a1df832fc4cb491d888fdf21182b5954f4

3 files changed, 8 insertions, 20 deletions

diff --git a/absl/base/call_once.h b/absl/base/call_once.h
index 5b468af8..96109f53 100644
--- a/absl/base/call_once.h
+++ b/absl/base/call_once.h
@@ -177,15 +177,8 @@ void CallOnceImpl(std::atomic<uint32_t>* control,
                                   scheduling_mode) == kOnceInit) {
     base_internal::invoke(std::forward<Callable>(fn),
                           std::forward<Args>(args)...);
-    // The call to SpinLockWake below is an optimization, because the waiter
-    // in SpinLockWait is waiting with a short timeout. The atomic load/store
-    // sequence is slightly faster than an atomic exchange:
-    //   old_control = control->exchange(base_internal::kOnceDone,
-    //                                   std::memory_order_release);
-    // We opt for a slightly faster case when there are no waiters, in spite
-    // of longer tail latency when there are waiters.
-    old_control = control->load(std::memory_order_relaxed);
-    control->store(base_internal::kOnceDone, std::memory_order_release);
+    old_control =
+        control->exchange(base_internal::kOnceDone, std::memory_order_release);
     if (old_control == base_internal::kOnceWaiter) {
       base_internal::SpinLockWake(control, true);
     }
diff --git a/absl/base/internal/spinlock.h b/absl/base/internal/spinlock.h
index e6ac9e64..dce1c854 100644
--- a/absl/base/internal/spinlock.h
+++ b/absl/base/internal/spinlock.h
@@ -15,11 +15,8 @@
 //
 
 //  Most users requiring mutual exclusion should use Mutex.
-//  SpinLock is provided for use in three situations:
+//  SpinLock is provided for use in two situations:
 //   - for use in code that Mutex itself depends on
-//   - to get a faster fast-path release under low contention (without an
-//     atomic read-modify-write) In return, SpinLock has worse behaviour under
-//     contention, which is why Mutex is preferred in most situations.
 //   - for async signal safety (see below)
 
 // SpinLock is async signal safe.  If a spinlock is used within a signal
diff --git a/absl/base/internal/spinlock_wait.h b/absl/base/internal/spinlock_wait.h
index 169bc749..c34ce41c 100644
--- a/absl/base/internal/spinlock_wait.h
+++ b/absl/base/internal/spinlock_wait.h
@@ -43,18 +43,16 @@ uint32_t SpinLockWait(std::atomic<uint32_t> *w, int n,
                       const SpinLockWaitTransition trans[],
                       SchedulingMode scheduling_mode);
 
-// If possible, wake some thread that has called SpinLockDelay(w, ...). If
-// "all" is true, wake all such threads.  This call is a hint, and on some
-// systems it may be a no-op; threads calling SpinLockDelay() will always wake
-// eventually even if SpinLockWake() is never called.
+// If possible, wake some thread that has called SpinLockDelay(w, ...). If `all`
+// is true, wake all such threads. On some systems, this may be a no-op; on
+// those systems, threads calling SpinLockDelay() will always wake eventually
+// even if SpinLockWake() is never called.
 void SpinLockWake(std::atomic<uint32_t> *w, bool all);
 
 // Wait for an appropriate spin delay on iteration "loop" of a
 // spin loop on location *w, whose previously observed value was "value".
 // SpinLockDelay() may do nothing, may yield the CPU, may sleep a clock tick,
-// or may wait for a delay that can be truncated by a call to SpinLockWake(w).
-// In all cases, it must return in bounded time even if SpinLockWake() is not
-// called.
+// or may wait for a call to SpinLockWake(w).
 void SpinLockDelay(std::atomic<uint32_t> *w, uint32_t value, int loop,
                    base_internal::SchedulingMode scheduling_mode);