Revert "Remove separate SSE4 unpremultiply function"

Could causedSSE4 instructions to be used on non SSE4 machines
in cases when qUnpremultiplywas not inlined.

This reverts commit 964ccc58534aac436529007000d1c38d76c88834.

Change-Id: Ic676ade8f75129e8d37c4d96cbfb2bdb5b794919
Task-number: QTBUG-45741
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

4 files changed, 34 insertions, 41 deletions

diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index 2817a4a145..37f2de98b6 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -6704,22 +6704,24 @@ void qInitDrawhelperAsm()
     }
 #endif // SSSE3
 
-#if defined(QT_COMPILER_SUPPORTS_SSE4_1) && !defined(__SSE4_1__)
+#if QT_COMPILER_SUPPORTS_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 defined(QT_COMPILER_SUPPORTS_AVX2) && !defined(__AVX2__)
+#if 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 8e0b2cbb18..43a3958997 100644
--- a/src/gui/painting/qdrawhelper_sse4.cpp
+++ b/src/gui/painting/qdrawhelper_sse4.cpp
@@ -32,6 +32,7 @@
 ****************************************************************************/
 
 #include <private/qdrawhelper_p.h>
+#include <private/qdrawingprimitive_sse2_p.h>
 
 #if defined(QT_COMPILER_SUPPORTS_SSE4_1)
 
@@ -53,7 +54,7 @@ const uint *QT_FASTCALL convertARGB32FromARGB32PM_sse4(uint *buffer, const uint
                                                        const QPixelLayout *, const QRgb *)
 {
     for (int i = 0; i < count; ++i)
-        buffer[i] = qUnpremultiply(src[i]);
+        buffer[i] = qUnpremultiply_sse4(src[i]);
     return buffer;
 }
 
@@ -61,7 +62,7 @@ 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(src[i]));
+        buffer[i] = ARGB2RGBA(qUnpremultiply_sse4(src[i]));
     return buffer;
 }
 
@@ -69,7 +70,7 @@ 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(src[i]));
+        buffer[i] = ARGB2RGBA(0xff000000 | qUnpremultiply_sse4(src[i]));
     return buffer;
 }
 
diff --git a/src/gui/painting/qdrawingprimitive_sse2_p.h b/src/gui/painting/qdrawingprimitive_sse2_p.h
index fbee06e157..4d0790a502 100644
--- a/src/gui/painting/qdrawingprimitive_sse2_p.h
+++ b/src/gui/painting/qdrawingprimitive_sse2_p.h
@@ -236,4 +236,25 @@ 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 05b3a76bce..f7f2185bef 100644
--- a/src/gui/painting/qrgb.h
+++ b/src/gui/painting/qrgb.h
@@ -36,11 +36,6 @@
 
 #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
 
@@ -92,45 +87,19 @@ 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 || alpha == 0)
+    if (alpha == 255)
         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