Optimize WindowManager::flush() with fast_dword_copy().

3 files changed, 9 insertions, 34 deletions

diff --git a/Widgets/GraphicsBitmap.h b/Widgets/GraphicsBitmap.h
index 64d68d2134..2ab8f92bdc 100644
--- a/Widgets/GraphicsBitmap.h
+++ b/Widgets/GraphicsBitmap.h
@@ -17,6 +17,7 @@ public:
     Size size() const { return m_size; }
     int width() const { return m_size.width(); }
     int height() const { return m_size.height(); }
+    size_t pitch() const { return m_pitch; }
 
 private:
     explicit GraphicsBitmap(const Size&);
diff --git a/Widgets/Painter.cpp b/Widgets/Painter.cpp
index 5ae921bd6f..fbd09f65e4 100644
--- a/Widgets/Painter.cpp
+++ b/Widgets/Painter.cpp
@@ -7,36 +7,6 @@
 
 #define DEBUG_WIDGET_UNDERDRAW
 
-ALWAYS_INLINE void fast_dword_copy(dword* dest, const dword* src, size_t count)
-{
-#ifdef SERENITY
-    asm volatile(
-        "rep movsl\n"
-        : "=S"(src), "=D"(dest)
-        : "S"(src), "D"(dest), "c"(count)
-        : "memory"
-    );
-#else
-    memcpy(dest, src, count * sizeof(dword));
-#endif
-}
-
-ALWAYS_INLINE void fast_dword_fill(dword* dest, dword value, size_t count)
-{
-#ifdef SERENITY
-    asm volatile(
-        "rep stosl\n"
-        : "=D"(dest)
-        : "D"(dest), "c"(count), "a"(value)
-        : "memory"
-    );
-#else
-    for (size_t i = 0; x <= count; ++x) {
-        dest[i] = value;
-    }
-#endif
-}
-
 Painter::Painter(GraphicsBitmap& bitmap)
 {
     m_font = &Font::defaultFont();
diff --git a/Widgets/WindowManager.cpp b/Widgets/WindowManager.cpp
index 41785387fa..3f6ec1ef5a 100644
--- a/Widgets/WindowManager.cpp
+++ b/Widgets/WindowManager.cpp
@@ -346,10 +346,14 @@ void WindowManager::flush(const Rect& a_rect)
 {
     auto rect = Rect::intersection(a_rect, m_screen_rect);
 
-    for (int y = rect.top(); y <= rect.bottom(); ++y) {
-        auto* front_scanline = m_front_bitmap->scanline(y);
-        auto* back_scanline = m_back_bitmap->scanline(y);
-        memcpy(front_scanline + rect.x(), back_scanline + rect.x(), rect.width() * sizeof(RGBA32));
+    RGBA32* front_ptr = m_front_bitmap->scanline(rect.y()) + rect.x();
+    const RGBA32* back_ptr = m_back_bitmap->scanline(rect.y()) + rect.x();
+    size_t pitch = m_back_bitmap->pitch();
+
+    for (int y = 0; y < rect.height(); ++y) {
+        fast_dword_copy(front_ptr, back_ptr, rect.width());
+        front_ptr = (RGBA32*)((byte*)front_ptr + pitch);
+        back_ptr = (const RGBA32*)((const byte*)back_ptr + pitch);
     }
 
     m_framebuffer.flush();