Make qYieldCpu() public API

Rewritten to be a bit simpler, added a few more yield/YieldProcessor
alternatives, added RISC-V support.

[ChangeLog][QtCore] Added qYieldCpu() function.

Fixes: QTBUG-103014
Change-Id: I53335f845a1345299031fffd176f59032e7400f5
Reviewed-by: Allan Sandfeld Jensen <[email protected]>

9 files changed, 134 insertions, 47 deletions

diff --git a/src/concurrent/qtconcurrentthreadengine.cpp b/src/concurrent/qtconcurrentthreadengine.cpp
index 03e018b0a29..ce02d0c2c95 100644
--- a/src/concurrent/qtconcurrentthreadengine.cpp
+++ b/src/concurrent/qtconcurrentthreadengine.cpp
@@ -3,8 +3,6 @@
 
 #include "qtconcurrentthreadengine.h"
 
-#include <QtCore/private/qsimd_p.h>
-
 #if !defined(QT_NO_CONCURRENT) || defined(Q_QDOC)
 
 QT_BEGIN_NAMESPACE
diff --git a/src/corelib/CMakeLists.txt b/src/corelib/CMakeLists.txt
index a7798a18f77..3333bc1a629 100644
--- a/src/corelib/CMakeLists.txt
+++ b/src/corelib/CMakeLists.txt
@@ -251,6 +251,7 @@ qt_internal_add_module(Core
         thread/qthreadstorage.h
         thread/qtsan_impl.h
         thread/qwaitcondition.h thread/qwaitcondition_p.h
+        thread/qyieldcpu.h
         time/qcalendar.cpp time/qcalendar.h
         time/qcalendarbackend_p.h
         time/qcalendarmath_p.h
diff --git a/src/corelib/global/qlogging.cpp b/src/corelib/global/qlogging.cpp
index d33f0937207..d4a22957787 100644
--- a/src/corelib/global/qlogging.cpp
+++ b/src/corelib/global/qlogging.cpp
@@ -14,7 +14,6 @@
 #include "qdebug.h"
 #include "qmutex.h"
 #include <QtCore/private/qlocking_p.h>
-#include <QtCore/private/qsimd_p.h>
 #include "qloggingcategory.h"
 #ifndef QT_BOOTSTRAPPED
 #include "qelapsedtimer.h"
diff --git a/src/corelib/global/qsimd_p.h b/src/corelib/global/qsimd_p.h
index 031ac78c607..55adcef6598 100644
--- a/src/corelib/global/qsimd_p.h
+++ b/src/corelib/global/qsimd_p.h
@@ -378,49 +378,6 @@ static inline uint64_t qCpuFeatures()
 #define qCpuHasFeature(feature)     (((qCompilerCpuFeatures & CpuFeature ## feature) == CpuFeature ## feature) \
                                      || ((qCpuFeatures() & CpuFeature ## feature) == CpuFeature ## feature))
 
-/*
-    Small wrapper around x86's PAUSE and ARM's YIELD instructions.
-
-    This is completely different from QThread::yieldCurrentThread(), which is
-    an OS-level operation that takes the whole thread off the CPU.
-
-    This is just preventing one SMT thread from filling a core's pipeline with
-    speculated further loop iterations (which need to be expensively flushed on
-    final success) when it could just give those pipeline slots to a second SMT
-    thread that can do something useful with the core, such as unblocking this
-    SMT thread :)
-
-    So, instead of
-
-        while (!condition)
-            ;
-
-    it's better to use
-
-        while (!condition)
-            qYieldCpu();
-*/
-static inline void qYieldCpu()
-{
-#if defined(Q_PROCESSOR_X86)
-    _mm_pause();
-#elif defined(Q_PROCESSOR_ARM) && Q_PROCESSOR_ARM >= 7 /* yield was added in ARMv7 */
-#  if __has_builtin(__builtin_arm_yield) /* e.g. Clang */
-    __builtin_arm_yield();
-#  elif defined(Q_OS_INTEGRITY) || defined(Q_CC_GNU_ONLY)
-    /*
-       - Integrity is missing the arm_acle.h header
-       - GCC doesn't have __yield() in arm_acle.h
-         https://stackoverflow.com/a/70076751/134841
-         https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105416
-    */
-    asm volatile("yield"); /* this works everywhere */
-#  else
-    __yield(); /* this is what should work everywhere */
-#  endif
-#endif
-}
-
 #ifdef __cplusplus
 } // extern "C"
 
diff --git a/src/corelib/thread/qatomic_cxx11.h b/src/corelib/thread/qatomic_cxx11.h
index 4f55a53ae3b..47a7bc9a106 100644
--- a/src/corelib/thread/qatomic_cxx11.h
+++ b/src/corelib/thread/qatomic_cxx11.h
@@ -6,6 +6,7 @@
 #define QATOMIC_CXX11_H
 
 #include <QtCore/qgenericatomic.h>
+#include <QtCore/qyieldcpu.h>
 #include <atomic>
 
 QT_BEGIN_NAMESPACE
diff --git a/src/corelib/thread/qfutureinterface.cpp b/src/corelib/thread/qfutureinterface.cpp
index de35089b2cc..3cf2566bd6c 100644
--- a/src/corelib/thread/qfutureinterface.cpp
+++ b/src/corelib/thread/qfutureinterface.cpp
@@ -9,7 +9,6 @@
 #include <QtCore/qcoreapplication.h>
 #include <QtCore/qthread.h>
 #include <QtCore/qvarlengtharray.h>
-#include <QtCore/private/qsimd_p.h> // for qYieldCpu()
 #include <private/qthreadpool_p.h>
 #include <private/qobject_p.h>
 
diff --git a/src/corelib/thread/qyieldcpu.h b/src/corelib/thread/qyieldcpu.h
new file mode 100644
index 00000000000..b54e54568d4
--- /dev/null
+++ b/src/corelib/thread/qyieldcpu.h
@@ -0,0 +1,66 @@
+// Copyright (C) 2023 The Qt Company Ltd.
+// Copyright (C) 2023 Intel Corporation.
+// SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only
+
+#ifndef QYIELDCPU_H
+#define QYIELDCPU_H
+
+#include <QtCore/qcompilerdetection.h>
+#include <QtCore/qprocessordetection.h>
+#include <QtCore/qtconfigmacros.h>
+
+#ifdef Q_CC_MSVC_ONLY
+// MSVC defines _YIELD_PROCESSOR() in <xatomic.h>, but as that is a private
+// header, we include the public ones
+#  ifdef __cplusplus
+#    include <atomic>
+extern "C"
+#  endif
+void _mm_pause(void);       // the compiler recognizes as intrinsic
+#endif
+
+QT_BEGIN_NAMESPACE
+
+#ifdef Q_CC_GNU
+__attribute__((artificial))
+#endif
+Q_ALWAYS_INLINE void qYieldCpu(void) Q_DECL_NOEXCEPT;
+
+void qYieldCpu(void)
+#ifdef __cplusplus
+    noexcept
+#endif
+{
+#if __has_builtin(__yield)
+    __yield();              // Generic
+#elif defined(_YIELD_PROCESSOR) && defined(Q_CC_MSVC)
+    _YIELD_PROCESSOR();     // Generic; MSVC's <atomic>
+
+#elif __has_builtin(__builtin_ia32_pause)
+    __builtin_ia32_pause();
+#elif defined(Q_PROCESSOR_X86) && defined(Q_CC_GNU)
+    // GCC < 10 didn't have __has_builtin()
+    __builtin_ia32_pause();
+#elif defined(Q_PROCESSOR_X86) && defined(Q_CC_MSVC)
+    _mm_pause();
+#elif defined(Q_PROCESSOR_X86)
+    asm("pause");           // hopefully asm() works in this compiler
+
+#elif __has_builtin(__builtin_arm_yield)
+    __builtin_arm_yield();
+#elif defined(Q_PROCESSOR_ARM) && Q_PROCESSOR_ARM >= 7
+    asm("yield");           // this works everywhere
+
+#elif __has_builtin(__builtin_riscv_pause)
+    __builtin_riscv_pause();        // Zihintpause extension
+#elif defined(Q_PROCESSOR_RISCV)
+    asm("fence w, 0");              // a.k.a. "pause"
+
+#elif defined(_YIELD_PROCESSOR) && defined(Q_CC_GHS)
+    _YIELD_PROCESSOR;       // Green Hills (INTEGRITY), but only on ARM
+#endif
+}
+
+QT_END_NAMESPACE
+
+#endif // QYIELDCPU_H
diff --git a/src/corelib/thread/qyieldcpu.qdoc b/src/corelib/thread/qyieldcpu.qdoc
new file mode 100644
index 00000000000..9a5f693cb4f
--- /dev/null
+++ b/src/corelib/thread/qyieldcpu.qdoc
@@ -0,0 +1,59 @@
+// Copyright (C) 2023 The Qt Company Ltd.
+// Copyright (C) 2023 Intel Corporation.
+// SPDX-License-Identifier: LicenseRef-Qt-Commercial OR LGPL-3.0-only OR GPL-2.0-only OR GPL-3.0-only
+
+/*!
+    \fn qYieldCpu()
+    \inmodule QtCore
+    \ingroup thread
+    \relates QAtomicInteger
+    \relatesalso QAtomicPointer
+    \since 6.7
+
+    Pauses the execution of the current thread for an unspecified time, using
+    hardware instructions, without de-scheduling this thread. This function is
+    meant to be used in high-throughput loops where the code expects another
+    thread to modify an atomic variable. This is completely different from
+    QThread::yieldCurrentThread(), which is an OS-level operation that may take
+    the whole thread off the CPU and allow other threads (possibly belonging to
+    other processes) to run.
+
+    So, instead of
+    \code
+        while (!condition)
+            ;
+    \endcode
+
+    one should write
+    \code
+        while (!condition)
+            qYieldCpu();
+    \endcode
+
+    This is useful both with and without hardware multithreading on the same
+    core. In the case of hardware threads, it serves to prevent further
+    speculative execution filling up the pipeline, which could starve the
+    sibling thread of resources. Across cores and higher levels of separation,
+    it allows the cache coherency protocol to allocate the cache line being
+    modified and inspected to the logical processor whose result this code is
+    expecting.
+
+    It is also recommended to loop around code that does not modify the global
+    variable, to avoid contention in exclusively obtaining the memory location.
+    Therefore, an atomic modification loop such as a spinlock acquisition
+    should be:
+
+    \code
+        while (true) {
+            while (!readOnlyCondition(atomic))
+                qYieldCpu();
+            if (modify(atomic))
+                break;
+        }
+    \endcode
+
+    On x86 processors and on RISC-V processors with the \c{Zihintpause}
+    extension, this will emit the \c PAUSE instruction, which is ignored on
+    processors that don't support it; on ARMv7 or later ARM processors, it will
+    emit the \c{YIELD} instruction.
+*/
diff --git a/tests/auto/corelib/global/qglobal/qglobal.c b/tests/auto/corelib/global/qglobal/qglobal.c
index abe6ec4fde0..8f9d8b52397 100644
--- a/tests/auto/corelib/global/qglobal/qglobal.c
+++ b/tests/auto/corelib/global/qglobal/qglobal.c
@@ -3,6 +3,7 @@
 
 #include <QtCore/qglobal.h>
 #include <QtCore/qtversion.h>
+#include <QtCore/qyieldcpu.h>
 
 #ifdef Q_COMPILER_THREAD_LOCAL
 #  include <threads.h>
@@ -62,6 +63,12 @@ const char *tst_qVersion()
 #endif
 }
 
+void tst_qYieldCpu(void) Q_DECL_NOEXCEPT;
+void tst_qYieldCpu(void)
+{
+    qYieldCpu();
+}
+
 /* Static assertion */
 Q_STATIC_ASSERT(true);
 Q_STATIC_ASSERT(1);