Remove separate SSE4 unpremultiply function

Merges the SSE4 specific unpremultiply with the normal version, and
adds a SSE2 fallback. There was no reason to split the two since
compile time options will ensure the right version is inlined.

Also adds short-cut for 0 and 255 values.

Change-Id: Ie5aa262f6964219fd3062d4a498f697cf79a4595
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

6 files changed, 45 insertions, 39 deletions

diff --git a/src/gui/image/qimage_conversions.cpp b/src/gui/image/qimage_conversions.cpp
index 28e3a48689..8cb886e09b 100644
--- a/src/gui/image/qimage_conversions.cpp
+++ b/src/gui/image/qimage_conversions.cpp
@@ -117,7 +117,7 @@ static const uint *QT_FASTCALL convertRGB32ToARGB32PM(uint *buffer, const uint *
     return buffer;
 }
 
-#ifdef QT_COMPILER_SUPPORTS_SSE4_1
+#if defined(QT_COMPILER_SUPPORTS_SSE4_1) && !defined(__SSE4_1__)
 extern const uint *QT_FASTCALL convertRGB32FromARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
 #endif
 
@@ -144,7 +144,7 @@ void convert_generic(QImageData *dest, const QImageData *src, Qt::ImageConversio
         if (src->format == QImage::Format_RGB32)
             convertToARGB32PM = convertRGB32ToARGB32PM;
         if (dest->format == QImage::Format_RGB32) {
-#ifdef QT_COMPILER_SUPPORTS_SSE4_1
+#if defined(QT_COMPILER_SUPPORTS_SSE4_1) && !defined(__SSE4_1__)
             if (qCpuHasFeature(SSE4_1))
                 convertFromARGB32PM = convertRGB32FromARGB32PM_sse4;
             else
@@ -193,7 +193,7 @@ bool convert_generic_inplace(QImageData *data, QImage::Format dst_format, Qt::Im
         if (data->format == QImage::Format_RGB32)
             convertToARGB32PM = convertRGB32ToARGB32PM;
         if (dst_format == QImage::Format_RGB32) {
-#ifdef QT_COMPILER_SUPPORTS_SSE4_1
+#if defined(QT_COMPILER_SUPPORTS_SSE4_1) && !defined(__SSE4_1__)
             if (qCpuHasFeature(SSE4_1))
                 convertFromARGB32PM = convertRGB32FromARGB32PM_sse4;
             else
diff --git a/src/gui/image/qimage_sse4.cpp b/src/gui/image/qimage_sse4.cpp
index 5fad4f572a..fb63f5bff9 100644
--- a/src/gui/image/qimage_sse4.cpp
+++ b/src/gui/image/qimage_sse4.cpp
@@ -33,7 +33,6 @@
 
 #include <qimage.h>
 #include <private/qdrawhelper_p.h>
-#include <private/qdrawingprimitive_sse2_p.h>
 #include <private/qimage_p.h>
 #include <private/qsimd_p.h>
 
@@ -45,7 +44,7 @@ const uint *QT_FASTCALL convertRGB32FromARGB32PM_sse4(uint *buffer, const uint *
                                                       const QPixelLayout *, const QRgb *)
 {
     for (int i = 0; i < count; ++i)
-        buffer[i] = 0xff000000 | qUnpremultiply_sse4(src[i]);
+        buffer[i] = 0xff000000 | qUnpremultiply(src[i]);
     return buffer;
 }
 
diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index 37f2de98b6..2817a4a145 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -6704,24 +6704,22 @@ void qInitDrawhelperAsm()
     }
 #endif // SSSE3
 
-#if QT_COMPILER_SUPPORTS_SSE4_1
+#if defined(QT_COMPILER_SUPPORTS_SSE4_1) && !defined(__SSE4_1__)
     if (qCpuHasFeature(SSE4_1)) {
-#if !defined(__SSE4_1__)
         extern const uint *QT_FASTCALL convertARGB32ToARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
         extern const uint *QT_FASTCALL convertRGBA8888ToARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
-        qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_sse4;
-        qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_sse4;
-#endif
         extern const uint *QT_FASTCALL convertARGB32FromARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
         extern const uint *QT_FASTCALL convertRGBA8888FromARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
         extern const uint *QT_FASTCALL convertRGBXFromARGB32PM_sse4(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
+        qPixelLayouts[QImage::Format_ARGB32].convertToARGB32PM = convertARGB32ToARGB32PM_sse4;
+        qPixelLayouts[QImage::Format_RGBA8888].convertToARGB32PM = convertRGBA8888ToARGB32PM_sse4;
         qPixelLayouts[QImage::Format_ARGB32].convertFromARGB32PM = convertARGB32FromARGB32PM_sse4;
         qPixelLayouts[QImage::Format_RGBA8888].convertFromARGB32PM = convertRGBA8888FromARGB32PM_sse4;
         qPixelLayouts[QImage::Format_RGBX8888].convertFromARGB32PM = convertRGBXFromARGB32PM_sse4;
     }
 #endif
 
-#if QT_COMPILER_SUPPORTS_AVX2 && !defined(__AVX2__)
+#if defined(QT_COMPILER_SUPPORTS_AVX2) && !defined(__AVX2__)
     if (qCpuHasFeature(AVX2)) {
         extern const uint *QT_FASTCALL convertARGB32ToARGB32PM_avx2(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
         extern const uint *QT_FASTCALL convertRGBA8888ToARGB32PM_avx2(uint *buffer, const uint *src, int count, const QPixelLayout *, const QRgb *);
diff --git a/src/gui/painting/qdrawhelper_sse4.cpp b/src/gui/painting/qdrawhelper_sse4.cpp
index 43a3958997..8e0b2cbb18 100644
--- a/src/gui/painting/qdrawhelper_sse4.cpp
+++ b/src/gui/painting/qdrawhelper_sse4.cpp
@@ -32,7 +32,6 @@
 ****************************************************************************/
 
 #include <private/qdrawhelper_p.h>
-#include <private/qdrawingprimitive_sse2_p.h>
 
 #if defined(QT_COMPILER_SUPPORTS_SSE4_1)
 
@@ -54,7 +53,7 @@ const uint *QT_FASTCALL convertARGB32FromARGB32PM_sse4(uint *buffer, const uint
                                                        const QPixelLayout *, const QRgb *)
 {
     for (int i = 0; i < count; ++i)
-        buffer[i] = qUnpremultiply_sse4(src[i]);
+        buffer[i] = qUnpremultiply(src[i]);
     return buffer;
 }
 
@@ -62,7 +61,7 @@ const uint *QT_FASTCALL convertRGBA8888FromARGB32PM_sse4(uint *buffer, const uin
                                                          const QPixelLayout *, const QRgb *)
 {
     for (int i = 0; i < count; ++i)
-        buffer[i] = ARGB2RGBA(qUnpremultiply_sse4(src[i]));
+        buffer[i] = ARGB2RGBA(qUnpremultiply(src[i]));
     return buffer;
 }
 
@@ -70,7 +69,7 @@ const uint *QT_FASTCALL convertRGBXFromARGB32PM_sse4(uint *buffer, const uint *s
                                                      const QPixelLayout *, const QRgb *)
 {
     for (int i = 0; i < count; ++i)
-        buffer[i] = ARGB2RGBA(0xff000000 | qUnpremultiply_sse4(src[i]));
+        buffer[i] = ARGB2RGBA(0xff000000 | qUnpremultiply(src[i]));
     return buffer;
 }
 
diff --git a/src/gui/painting/qdrawingprimitive_sse2_p.h b/src/gui/painting/qdrawingprimitive_sse2_p.h
index 4d0790a502..fbee06e157 100644
--- a/src/gui/painting/qdrawingprimitive_sse2_p.h
+++ b/src/gui/painting/qdrawingprimitive_sse2_p.h
@@ -236,25 +236,4 @@ QT_END_NAMESPACE
 
 #endif // __SSE2__
 
-QT_BEGIN_NAMESPACE
-#if QT_COMPILER_SUPPORTS_HERE(SSE4_1)
-QT_FUNCTION_TARGET(SSE4_1)
-inline QRgb qUnpremultiply_sse4(QRgb p)
-{
-    const uint alpha = qAlpha(p);
-    const uint invAlpha = qt_inv_premul_factor[alpha];
-    const __m128i via = _mm_set1_epi32(invAlpha);
-    const __m128i vr = _mm_set1_epi32(0x8000);
-    __m128i vl = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(p));
-    vl = _mm_mullo_epi32(vl, via);
-    vl = _mm_add_epi32(vl, vr);
-    vl = _mm_srai_epi32(vl, 16);
-    vl = _mm_insert_epi32(vl, alpha, 3);
-    vl = _mm_packus_epi32(vl, _mm_setzero_si128());
-    vl = _mm_packus_epi16(vl, _mm_setzero_si128());
-    return _mm_cvtsi128_si32(vl);
-}
-#endif
-QT_END_NAMESPACE
-
 #endif // QDRAWINGPRIMITIVE_SSE2_P_H
diff --git a/src/gui/painting/qrgb.h b/src/gui/painting/qrgb.h
index f7f2185bef..05b3a76bce 100644
--- a/src/gui/painting/qrgb.h
+++ b/src/gui/painting/qrgb.h
@@ -36,6 +36,11 @@
 
 #include <QtCore/qglobal.h>
 #include <QtCore/qprocessordetection.h>
+#if defined(__SSE4_1__)
+#include <smmintrin.h>
+#elif defined(__SSE2__)
+#include <emmintrin.h>
+#endif
 
 QT_BEGIN_NAMESPACE
 
@@ -87,19 +92,45 @@ inline Q_DECL_RELAXED_CONSTEXPR QRgb qPremultiply(QRgb x)
 
 Q_GUI_EXPORT extern const uint qt_inv_premul_factor[];
 
+#if defined(__SSE2__)
+inline QRgb qUnpremultiply(QRgb p)
+{
+    const uint alpha = qAlpha(p);
+    if (alpha == 255 || alpha == 0)
+        return p;
+    const uint invAlpha = qt_inv_premul_factor[alpha];
+    const __m128i via = _mm_set1_epi32(invAlpha);
+    const __m128i vr = _mm_set1_epi32(0x8000);
+#ifdef __SSE4_1__
+    __m128i vl = _mm_cvtepu8_epi32(_mm_cvtsi32_si128(p));
+    vl = _mm_mullo_epi32(vl, via);
+#else
+    __m128i vl = _mm_unpacklo_epi8(_mm_cvtsi32_si128(p), _mm_setzero_si128());
+    vl = _mm_unpacklo_epi16(vl, vl);
+    __m128i vll = _mm_mullo_epi16(vl, via);
+    __m128i vlh = _mm_mulhi_epu16(vl, via);
+    vl = _mm_add_epi32(vll, _mm_slli_epi32(vlh, 16));
+#endif
+    vl = _mm_add_epi32(vl, vr);
+    vl = _mm_srli_epi32(vl, 16);
+    vl = _mm_packs_epi32(vl, _mm_setzero_si128());
+    vl = _mm_insert_epi16(vl, alpha, 3);
+    vl = _mm_packus_epi16(vl, _mm_setzero_si128());
+    return _mm_cvtsi128_si32(vl);
+}
+#else
 inline QRgb qUnpremultiply(QRgb p)
 {
     const uint alpha = qAlpha(p);
     // Alpha 255 and 0 are the two most common values, which makes them beneficial to short-cut.
-    if (alpha == 255)
+    if (alpha == 255 || alpha == 0)
         return p;
-    if (alpha == 0)
-        return 0;
     // (p*(0x00ff00ff/alpha)) >> 16 == (p*255)/alpha for all p and alpha <= 256.
     const uint invAlpha = qt_inv_premul_factor[alpha];
     // We add 0x8000 to get even rounding. The rounding also ensures that qPremultiply(qUnpremultiply(p)) == p for all p.
     return qRgba((qRed(p)*invAlpha + 0x8000)>>16, (qGreen(p)*invAlpha + 0x8000)>>16, (qBlue(p)*invAlpha + 0x8000)>>16, alpha);
 }
+#endif
 
 QT_END_NAMESPACE