Replace the SSE prologues by a macro

Replace the code of the SSE prologue by a macro to avoid copying the
prologue everywhere.

Reviewed-by: Andreas Kling

3 files changed, 9 insertions, 17 deletions

diff --git a/src/corelib/tools/qsimd_p.h b/src/corelib/tools/qsimd_p.h
index 5ff0f97..a3148fb 100644
--- a/src/corelib/tools/qsimd_p.h
+++ b/src/corelib/tools/qsimd_p.h
@@ -150,7 +150,9 @@ enum CPUFeatures {
 
 Q_CORE_EXPORT uint qDetectCPUFeatures();
 
-Q_CORE_EXPORT uint qDetectCPUFeatures();
+
+#define ALIGNMENT_PROLOGUE_16BYTES(ptr, i, length) \
+    for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((4 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x3)) & 0x3))); ++i)
 
 QT_END_NAMESPACE
 
diff --git a/src/gui/painting/qdrawhelper_sse2.cpp b/src/gui/painting/qdrawhelper_sse2.cpp
index 7ab9eda..22c0384 100644
--- a/src/gui/painting/qdrawhelper_sse2.cpp
+++ b/src/gui/painting/qdrawhelper_sse2.cpp
@@ -112,9 +112,7 @@ void qt_blend_rgb32_on_rgb32_sse2(uchar *destPixels, int dbpl,
                 int x = 0;
 
                 // First, align dest to 16 bytes:
-                const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;
-                const int prologLength = qMin(w, offsetToAlignOn16Bytes);
-                for (; x < prologLength; ++x) {
+                ALIGNMENT_PROLOGUE_16BYTES(dst, x, w) {
                     quint32 s = src[x];
                     s = BYTE_MUL(s, const_alpha);
                     dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
@@ -182,12 +180,10 @@ inline int comp_func_Plus_one_pixel(uint d, const uint s)
 void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uint const_alpha)
 {
     int x = 0;
-    const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;
-    const int prologLength = qMin(length, offsetToAlignOn16Bytes);
 
     if (const_alpha == 255) {
         // 1) Prologue: align destination on 16 bytes
-        for (; x < prologLength; ++x)
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
             dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
 
         // 2) composition with SSE2
@@ -208,7 +204,7 @@ void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uin
         const __m128i oneMinusConstAlpha =  _mm_set1_epi16(one_minus_const_alpha);
 
         // 1) Prologue: align destination on 16 bytes
-        for (; x < prologLength; ++x)
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
             dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
 
         const __m128i half = _mm_set1_epi16(0x80);
@@ -239,9 +235,7 @@ void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, u
         int x = 0;
 
         // 1) prologue, align on 16 bytes
-        const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;
-        const int prologLength = qMin(length, offsetToAlignOn16Bytes);
-        for (; x < prologLength; ++x)
+        ALIGNMENT_PROLOGUE_16BYTES(dst, x, length)
             dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
 
         // 2) interpolate pixels with SSE2
diff --git a/src/gui/painting/qdrawingprimitive_sse2_p.h b/src/gui/painting/qdrawingprimitive_sse2_p.h
index 18355c2..d8f6bf5 100644
--- a/src/gui/painting/qdrawingprimitive_sse2_p.h
+++ b/src/gui/painting/qdrawingprimitive_sse2_p.h
@@ -143,9 +143,7 @@ QT_BEGIN_NAMESPACE
     int x = 0; \
 \
     /* First, get dst aligned. */ \
-    const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;\
-    const int prologLength = qMin(length, offsetToAlignOn16Bytes);\
-    for (; x < prologLength; ++x) { \
+    ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \
         uint s = src[x]; \
         if (s >= 0xff000000) \
             dst[x] = s; \
@@ -202,9 +200,7 @@ QT_BEGIN_NAMESPACE
 { \
     int x = 0; \
 \
-    const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;\
-    const int prologLength = qMin(length, offsetToAlignOn16Bytes);\
-    for (; x < prologLength; ++x) { \
+    ALIGNMENT_PROLOGUE_16BYTES(dst, x, length) { \
         quint32 s = src[x]; \
         if (s != 0) { \
             s = BYTE_MUL(s, const_alpha); \