From 4a43149d7b1ba10a96826a2bb38f1d5d608b6cb0 Mon Sep 17 00:00:00 2001
From: Thiago Macieira <thiago.macieira@intel.com>
Date: Fri, 7 Jul 2023 11:16:00 -0700
Subject: [PATCH] QThreadPipe & QEvent: use a simpler atomic operation

This causes modern compilers (i.e., not MSVC) to emit a single
bit-test-and-set instruction instead of a cmpxchg. It's still an atomic
operation, so it's not that much faster (if at all), but it has simpler
encoding.

Previous:
000000000026bca0 <QEventDispatcherUNIX::wakeUp()>:
  26bca0:       mov    0x8(%rdi),%rdx
  26bca4:       xor    %eax,%eax
  26bca6:       mov    $0x1,%ecx
  26bcab:       lock cmpxchg %ecx,0x98(%rdx)
  26bcb3:       jne    26bcc5 <QEventDispatcherUNIX::wakeUp()+0x25>
  26bcb5:       mov    0x90(%rdx),%edi
  26bcbb:       mov    $0x1,%esi
  26bcc0:       jmp    c01d0 <eventfd_write@plt>
  26bcc5:       ret

Now:
  26b3a0:       mov    0x8(%rdi),%rax
  26b3a4:       lock btsl $0x0,0x98(%rax)
  26b3ad:       jae    26b3b0 <QEventDispatcherUNIX::wakeUp()+0x10>
  26b3af:       ret
  26b3b0:       mov    0x90(%rax),%edi
  26b3b6:       mov    $0x1,%esi
  26b3bb:       jmp    c11d0 <eventfd_write@plt>

Change-Id: I53335f845a1345299031fffd176fa8ac1de3ad13
Reviewed-by: Fabian Kosmale <fabian.kosmale@qt.io>
---
 src/corelib/kernel/qcoreevent.cpp            | 5 +++--
 src/corelib/kernel/qeventdispatcher_unix.cpp | 2 +-
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/src/corelib/kernel/qcoreevent.cpp b/src/corelib/kernel/qcoreevent.cpp
index 46398872f7..838c90fce7 100644
--- a/src/corelib/kernel/qcoreevent.cpp
+++ b/src/corelib/kernel/qcoreevent.cpp
@@ -433,8 +433,9 @@ struct QBasicAtomicBitField {
         QBasicAtomicInteger<uint> &entry = data[which / BitsPerInt];
         const uint old = entry.loadRelaxed();
         const uint bit = 1U << (which % BitsPerInt);
-        return !(old & bit) // wasn't taken
-            && entry.testAndSetRelaxed(old, old | bit); // still wasn't taken
+        if (old & bit)
+            return false;       // already taken
+        return (entry.fetchAndOrRelaxed(bit) & bit) == 0;
 
         // don't update 'next' here - it's unlikely that it will need
         // to be updated, in the general case, and having 'next'
diff --git a/src/corelib/kernel/qeventdispatcher_unix.cpp b/src/corelib/kernel/qeventdispatcher_unix.cpp
index dc87a9cd63..12603d77f0 100644
--- a/src/corelib/kernel/qeventdispatcher_unix.cpp
+++ b/src/corelib/kernel/qeventdispatcher_unix.cpp
@@ -123,7 +123,7 @@ pollfd QThreadPipe::prepare() const
 
 void QThreadPipe::wakeUp()
 {
-    if (wakeUps.testAndSetAcquire(0, 1)) {
+    if ((wakeUps.fetchAndOrAcquire(1) & 1) == 0) {
 #if QT_CONFIG(eventfd)
         eventfd_write(fds[0], 1);
         return;