Make multi-threaded image transforms and painter fills configurable

Some users prefer to avoid having this many threads. This also
moves disabling it for WASM from sources to config.

Fixes: QTBUG-129650
Change-Id: Ib4c7903e85ba9cb75a9e013d1032653ea0ab8b84
Reviewed-by: Giuseppe D'Angelo <giuseppe.dangelo@kdab.com>

src/corelib/kernel/qcoreapplication.cpp

@@ -930,10 +930,8 @@ QCoreApplication::~QCoreApplication()
 #if QT_CONFIG(thread)
     // Synchronize and stop the global thread pool threads.
     QThreadPool *globalThreadPool = nullptr;
-    QThreadPool *guiThreadPool = nullptr;
     QT_TRY {
         globalThreadPool = QThreadPool::globalInstance();
-        guiThreadPool = QThreadPoolPrivate::qtGuiInstance();
     } QT_CATCH (...) {
         // swallow the exception, since destructors shouldn't throw
     }
@@ -941,10 +939,6 @@ QCoreApplication::~QCoreApplication()
         globalThreadPool->waitForDone();
         delete globalThreadPool;
     }
-    if (guiThreadPool) {
-        guiThreadPool->waitForDone();
-        delete guiThreadPool;
-    }
 #endif
 
 #ifndef QT_NO_QOBJECT

src/corelib/thread/qthreadpool.cpp

@@ -474,28 +474,6 @@ QThreadPool *QThreadPool::globalInstance()
     return theInstance;
 }
 
-/*!
-    Returns the QThreadPool instance for Qt Gui.
-    \internal
-*/
-QThreadPool *QThreadPoolPrivate::qtGuiInstance()
-{
-    Q_CONSTINIT static QPointer<QThreadPool> guiInstance;
-    Q_CONSTINIT static QBasicMutex theMutex;
-    const static bool runtime_disable = qEnvironmentVariableIsSet("QT_NO_GUI_THREADPOOL");
-    if (runtime_disable)
-        return nullptr;
-    const QMutexLocker locker(&theMutex);
-    if (guiInstance.isNull() && !QCoreApplication::closingDown()) {
-        guiInstance = new QThreadPool();
-        // Limit max thread to avoid too many parallel threads.
-        // We are not optimized for much more than 4 or 8 threads.
-        if (guiInstance && guiInstance->maxThreadCount() > 4)
-            guiInstance->setMaxThreadCount(qBound(4, guiInstance->maxThreadCount() / 2, 8));
-    }
-    return guiInstance;
-}
-
 /*!
     Reserves a thread and uses it to run \a runnable, unless this thread will
     make the current thread count exceed maxThreadCount().  In that case,

src/corelib/thread/qthreadpool_p.h

@@ -133,8 +133,6 @@ public:
     void stealAndRunRunnable(QRunnable *runnable);
     void deletePageIfFinished(QueuePage *page);
 
-    static QThreadPool *qtGuiInstance();
-
     mutable QMutex mutex;
     QSet<QThreadPoolThread *> allThreads;
     QQueue<QThreadPoolThread *> waitingThreads;

src/gui/configure.cmake

@@ -1271,6 +1271,14 @@ qt_feature("raster-fp" PRIVATE
     PURPOSE "Internal painting support for floating point rasterization."
     CONDITION NOT VXWORKS # QTBUG-115777
 )
+
+qt_feature("qtgui-threadpool" PRIVATE
+    SECTION "Painting"
+    LABEL "Multi-threaded image and painting helpers"
+    PURPOSE "Multi-threaded image transforms and QPainter fills."
+    CONDITION QT_FEATURE_thread AND NOT WASM
+)
+
 qt_feature("undocommand" PUBLIC
     SECTION "Utilities"
     LABEL "QUndoCommand"
@@ -1344,6 +1352,10 @@ qt_configure_add_summary_entry(ARGS "vulkan")
 qt_configure_add_summary_entry(ARGS "metal")
 qt_configure_add_summary_entry(ARGS "graphicsframecapture")
 qt_configure_add_summary_entry(ARGS "sessionmanager")
+qt_configure_add_summary_entry(
+    ARGS "qtgui-threadpool"
+    CONDITION QT_FEATURE_thread
+)
 qt_configure_end_summary_section() # end of "Qt Gui" section
 qt_configure_add_summary_section(NAME "Features used by QPA backends")
 qt_configure_add_summary_entry(ARGS "evdev")

src/gui/image/qimage.cpp

@@ -35,7 +35,7 @@
 #include <private/qimage_p.h>
 #include <private/qfont_p.h>
 
-#if QT_CONFIG(thread)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
 #include <qthreadpool.h>
 #include <private/qthreadpool_p.h>
@@ -4908,7 +4908,7 @@ QImage Q_TRACE_INSTRUMENT(qtgui) QImage::transformed(const QTransform &matrix, Q
         }
         // Otherwise only use it when the scaling factor demands it, or the image is large enough to scale multi-threaded
         if (nonpaintable_scale_xform
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
             || (ws * hs) >= (1<<20)
 #endif
             ) {
@@ -5328,10 +5328,10 @@ void QImage::applyColorTransform(const QColorTransform &transform)
         };
     }
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(width()) * height()) >> 16;
     segments = std::min(segments, height());
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;
@@ -5818,10 +5818,10 @@ QImage QImage::colorTransformed(const QColorTransform &transform, QImage::Format
         }
     }
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(width()) * height()) >> 16;
     segments = std::min(segments, height());
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;

src/gui/image/qimage_conversions.cpp

@@ -17,11 +17,6 @@
 #include <qsemaphore.h>
 #include <qthreadpool.h>
 #include <private/qthreadpool_p.h>
-#ifdef Q_OS_WASM
-// WebAssembly has threads; however we can't block the main thread.
-#else
-#define QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
-#endif
 #endif
 
 #include <QtCore/q20utility.h>
@@ -212,11 +207,11 @@ void convert_generic(QImageData *dest, const QImageData *src, Qt::ImageConversio
         }
     };
 
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(src->width) * src->height) >> 16;
     segments = std::min(segments, src->height);
 
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments <= 1 || !threadPool || threadPool->contains(QThread::currentThread()))
         return convertSegment(0, src->height);
 
@@ -267,11 +262,11 @@ void convert_generic_over_rgb64(QImageData *dest, const QImageData *src, Qt::Ima
             destData += dest->bytes_per_line;
         }
     };
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(src->width) * src->height) >> 16;
     segments = std::min(segments, src->height);
 
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments <= 1 || !threadPool || threadPool->contains(QThread::currentThread()))
         return convertSegment(0, src->height);
 
@@ -321,11 +316,11 @@ void convert_generic_over_rgba32f(QImageData *dest, const QImageData *src, Qt::I
             destData += dest->bytes_per_line;
         }
     };
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(src->width) * src->height) >> 16;
     segments = std::min(segments, src->height);
 
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments <= 1 || !threadPool || threadPool->contains(QThread::currentThread()))
         return convertSegment(0, src->height);
 
@@ -434,10 +429,10 @@ bool convert_generic_inplace(QImageData *data, QImage::Format dst_format, Qt::Im
             destData += params.bytesPerLine;
         }
     };
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(data->width) * data->height) >> 16;
     segments = std::min(segments, data->height);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;
@@ -527,10 +522,10 @@ bool convert_generic_inplace_over_rgb64(QImageData *data, QImage::Format dst_for
             destData += params.bytesPerLine;
         }
     };
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(data->width) * data->height) >> 16;
     segments = std::min(segments, data->height);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;
@@ -621,10 +616,10 @@ bool convert_generic_inplace_over_rgba32f(QImageData *data, QImage::Format dst_f
             destData += params.bytesPerLine;
         }
     };
-#ifdef QT_USE_THREAD_PARALLEL_IMAGE_CONVERSIONS
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(data->width) * data->height) >> 16;
     segments = std::min(segments, data->height);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;

src/gui/kernel/qguiapplication.cpp

@@ -99,6 +99,10 @@
 #include <private/qvulkandefaultinstance_p.h>
 #endif
 
+#if QT_CONFIG(thread)
+#include <QtCore/QThreadPool>
+#endif
+
 #include <qtgui_tracepoints_p.h>
 
 #include <private/qtools_p.h>
@@ -686,6 +690,20 @@ QGuiApplication::~QGuiApplication()
     d->cursor_list.clear();
 #endif
 
+#if QT_CONFIG(qtgui_threadpool)
+    // Synchronize and stop the gui thread pool threads.
+    QThreadPool *guiThreadPool = nullptr;
+    QT_TRY {
+        guiThreadPool = QGuiApplicationPrivate::qtGuiThreadPool();
+    } QT_CATCH (...) {
+        // swallow the exception, since destructors shouldn't throw
+    }
+    if (guiThreadPool) {
+        guiThreadPool->waitForDone();
+        delete guiThreadPool;
+    }
+#endif
+
     delete QGuiApplicationPrivate::app_icon;
     QGuiApplicationPrivate::app_icon = nullptr;
     delete QGuiApplicationPrivate::platform_name;
@@ -4552,6 +4570,32 @@ QInputDeviceManager *QGuiApplicationPrivate::inputDeviceManager()
     return m_inputDeviceManager;
 }
 
+/*!
+    Returns the QThreadPool instance for Qt Gui.
+    \internal
+*/
+QThreadPool *QGuiApplicationPrivate::qtGuiThreadPool()
+{
+#if QT_CONFIG(qtgui_threadpool)
+    Q_CONSTINIT static QPointer<QThreadPool> guiInstance;
+    Q_CONSTINIT static QBasicMutex theMutex;
+    const static bool runtime_disable = qEnvironmentVariableIsSet("QT_NO_GUI_THREADPOOL");
+    if (runtime_disable)
+        return nullptr;
+    const QMutexLocker locker(&theMutex);
+    if (guiInstance.isNull() && !QCoreApplication::closingDown()) {
+        guiInstance = new QThreadPool();
+        // Limit max thread to avoid too many parallel threads.
+        // We are not optimized for much more than 4 or 8 threads.
+        if (guiInstance && guiInstance->maxThreadCount() > 4)
+            guiInstance->setMaxThreadCount(qBound(4, guiInstance->maxThreadCount() / 2, 8));
+    }
+    return guiInstance;
+#else
+    return nullptr;
+#endif
+}
+
 /*!
     \fn template <typename QNativeInterface> QNativeInterface *QGuiApplication::nativeInterface() const
 

src/gui/kernel/qguiapplication_p.h

@@ -57,6 +57,7 @@ class QActionPrivate;
 #if QT_CONFIG(shortcut)
 class QShortcutPrivate;
 #endif
+class QThreadPool;
 
 class Q_GUI_EXPORT QGuiApplicationPrivate : public QCoreApplicationPrivate
 {
@@ -337,6 +338,8 @@ public:
 
     static QEvent::Type contextMenuEventType();
 
+    static QThreadPool *qtGuiThreadPool();
+
 protected:
     virtual void handleThemeChanged();
 

src/gui/painting/qdrawhelper.cpp

@@ -27,11 +27,7 @@
 #include <qloggingcategory.h>
 #include <qmath.h>
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
-#define QT_USE_THREAD_PARALLEL_FILLS
-#endif
-
-#if defined(QT_USE_THREAD_PARALLEL_FILLS)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
 #include <qthreadpool.h>
 #include <private/qthreadpool_p.h>
@@ -3963,10 +3959,10 @@ static void spanfill_from_first(QRasterBuffer *rasterBuffer, QPixelLayout::BPP b
 
 // -------------------- blend methods ---------------------
 
-#if defined(QT_USE_THREAD_PARALLEL_FILLS)
+#if QT_CONFIG(qtgui_threadpool)
 #define QT_THREAD_PARALLEL_FILLS(function) \
     const int segments = (count + 32) / 64; \
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance(); \
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool(); \
     if (segments > 1 && qPixelLayouts[data->rasterBuffer->format].bpp >= QPixelLayout::BPP8 \
              && threadPool && !threadPool->contains(QThread::currentThread())) { \
         QSemaphore semaphore; \

src/gui/painting/qimagescale.cpp

@@ -10,9 +10,10 @@
 #include "qrgba64_p.h"
 #include "qrgbafloat.h"
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
 #include <qthreadpool.h>
+#include <private/qguiapplication_p.h>
 #include <private/qthreadpool_p.h>
 #endif
 
@@ -284,10 +285,10 @@ void qt_qimageScaleAARGBA_down_xy_neon(QImageScaleInfo *isi, unsigned int *dest,
 template<typename T>
 static inline void multithread_pixels_function(QImageScaleInfo *isi, int dh, const T &scaleSection)
 {
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(isi->sh) * isi->sw) / (1<<16);
     segments = std::min(segments, dh);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;

src/gui/painting/qimagescale_lsx.cpp

@@ -6,8 +6,9 @@
 #include <private/qdrawhelper_loongarch64_p.h>
 #include <private/qsimd_p.h>
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
+#include <private/qguiapplication_p.h>
 #include <private/qthreadpool_p.h>
 #endif
 
@@ -20,10 +21,10 @@ using namespace QImageScale;
 template<typename T>
 static inline void multithread_pixels_function(QImageScaleInfo *isi, int dh, const T &scaleSection)
 {
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(isi->sh) * isi->sw) / (1<<16);
     segments = std::min(segments, dh);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;

src/gui/painting/qimagescale_neon.cpp

@@ -3,11 +3,13 @@
 
 #include "qimagescale_p.h"
 #include "qimage.h"
+#include <private/qtguiglobal_p.h>
 #include <private/qsimd_p.h>
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
 #include <qthreadpool.h>
+#include <private/qguiapplication_p.h>
 #include <private/qthreadpool_p.h>
 #endif
 
@@ -20,10 +22,10 @@ using namespace QImageScale;
 template<typename T>
 static inline void multithread_pixels_function(QImageScaleInfo *isi, int dh, const T &scaleSection)
 {
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(isi->sh) * isi->sw) / (1<<16);
     segments = std::min(segments, dh);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;

src/gui/painting/qimagescale_sse4.cpp

@@ -6,9 +6,10 @@
 #include <private/qdrawhelper_x86_p.h>
 #include <private/qsimd_p.h>
 
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
 #include <qsemaphore.h>
 #include <qthreadpool.h>
+#include <private/qguiapplication_p.h>
 #include <private/qthreadpool_p.h>
 #endif
 
@@ -21,10 +22,10 @@ using namespace QImageScale;
 template<typename T>
 static inline void multithread_pixels_function(QImageScaleInfo *isi, int dh, const T &scaleSection)
 {
-#if QT_CONFIG(thread) && !defined(Q_OS_WASM)
+#if QT_CONFIG(qtgui_threadpool)
     int segments = (qsizetype(isi->sh) * isi->sw) / (1<<16);
     segments = std::min(segments, dh);
-    QThreadPool *threadPool = QThreadPoolPrivate::qtGuiInstance();
+    QThreadPool *threadPool = QGuiApplicationPrivate::qtGuiThreadPool();
     if (segments > 1 && threadPool && !threadPool->contains(QThread::currentThread())) {
         QSemaphore semaphore;
         int y = 0;