Optimized fetchUntransformed RGB888

Reuses the optimized routines from qimage to make painting RGB888
images faster on SSSE3 and NEON.

Change-Id: I99116b318322ba4cb0ddc2cb90bcf17a0350ef99
Reviewed-by: Gunnar Sletta <gunnar@sletta.org>

src/gui/painting/qdrawhelper.cpp

@@ -6382,11 +6382,15 @@ static void qInitDrawhelperFunctions()
                                                     int const_alpha);
 
         extern void QT_FASTCALL storePixelsBPP24_ssse3(uchar *dest, const uint *src, int index, int count);
+        extern const uint * QT_FASTCALL qt_fetchUntransformed_888_ssse3(uint *buffer, const Operator *, const QSpanData *data,
+                                                                        int y, int x, int length);
         qBlendFunctions[QImage::Format_RGB32][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_ssse3;
         qBlendFunctions[QImage::Format_ARGB32_Premultiplied][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_argb32_ssse3;
         qBlendFunctions[QImage::Format_RGBX8888][QImage::Format_RGBA8888_Premultiplied] = qt_blend_argb32_on_argb32_ssse3;
         qBlendFunctions[QImage::Format_RGBA8888_Premultiplied][QImage::Format_RGBA8888_Premultiplied] = qt_blend_argb32_on_argb32_ssse3;
         qStorePixels[QPixelLayout::BPP24] = storePixelsBPP24_ssse3;
+        sourceFetch[BlendUntransformed][QImage::Format_RGB888] = qt_fetchUntransformed_888_ssse3;
+        sourceFetch[BlendTiled][QImage::Format_RGB888] = qt_fetchUntransformed_888_ssse3;
     }
 #endif // SSSE3
 
@@ -6449,6 +6453,9 @@ static void qInitDrawhelperFunctions()
 
     qt_fetch_radial_gradient = qt_fetch_radial_gradient_neon;
 
+    sourceFetch[BlendUntransformed][QImage::Format_RGB888] = qt_fetchUntransformed_888_neon;
+    sourceFetch[BlendTiled][QImage::Format_RGB888] = qt_fetchUntransformed_888_neon;
+
 #if defined(ENABLE_PIXMAN_DRAWHELPERS)
     // The RGB16 helpers are using Arm32 assemblythat has not been ported to AArch64
     qBlendFunctions[QImage::Format_RGB16][QImage::Format_ARGB32_Premultiplied] = qt_blend_argb32_on_rgb16_neon;

src/gui/painting/qdrawhelper_neon.cpp

@@ -1061,6 +1061,16 @@ const uint * QT_FASTCALL qt_fetch_radial_gradient_neon(uint *buffer, const Opera
     return qt_fetch_radial_gradient_template<QRadialFetchSimd<QSimdNeon>,uint>(buffer, op, data, y, x, length);
 }
 
+extern void QT_FASTCALL qt_convert_rgb888_to_rgb32_neon(quint32 *dst, const uchar *src, int len);
+
+const uint * QT_FASTCALL qt_fetchUntransformed_888_neon(uint *buffer, const Operator *, const QSpanData *data,
+                                                       int y, int x, int length)
+{
+    const uchar *line = data->texture.scanLine(y) + x * 3;
+    qt_convert_rgb888_to_rgb32_neon(buffer, line, length);
+    return buffer;
+}
+
 QT_END_NAMESPACE
 
 #endif // __ARM_NEON__

src/gui/painting/qdrawhelper_neon_p.h

@@ -137,6 +137,9 @@ void QT_FASTCALL qt_destStoreRGB16_neon(QRasterBuffer *rasterBuffer,
 void QT_FASTCALL comp_func_solid_SourceOver_neon(uint *destPixels, int length, uint color, uint const_alpha);
 void QT_FASTCALL comp_func_Plus_neon(uint *dst, const uint *src, int length, uint const_alpha);
 
+const uint * QT_FASTCALL qt_fetchUntransformed_888_neon(uint *buffer, const Operator *, const QSpanData *data,
+                                                       int y, int x, int length);
+
 #endif // __ARM_NEON__
 
 QT_END_NAMESPACE

src/gui/painting/qdrawhelper_ssse3.cpp

@@ -233,6 +233,16 @@ void QT_FASTCALL storePixelsBPP24_ssse3(uchar *dest, const uint *src, int index,
     store_uint24_ssse3(dest + index * 3, src, count);
 }
 
+extern void QT_FASTCALL qt_convert_rgb888_to_rgb32_ssse3(quint32 *dst, const uchar *src, int len);
+
+const uint * QT_FASTCALL qt_fetchUntransformed_888_ssse3(uint *buffer, const Operator *, const QSpanData *data,
+                                                         int y, int x, int length)
+{
+    const uchar *line = data->texture.scanLine(y) + x * 3;
+    qt_convert_rgb888_to_rgb32_ssse3(buffer, line, length);
+    return buffer;
+}
+
 QT_END_NAMESPACE
 
 #endif // QT_COMPILER_SUPPORTS_SSSE3