Add SSE2 qt_memfill64

Implemented by merging with the qt_memfill32 implementation in a
non-inlining function.

Change-Id: I343f2beed55440a7ac0bfffd15636f8ba995a2bd
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>

src/gui/painting/qdrawhelper.cpp

@@ -6260,10 +6260,12 @@ DrawHelper qDrawHelper[QImage::NImageFormats] =
     },
 };
 
+#if !defined(__SSE2__)
 void qt_memfill64(quint64 *dest, quint64 color, qsizetype count)
 {
     qt_memfill_template<quint64>(dest, color, count);
 }
+#endif
 
 void qt_memfill16(quint16 *dest, quint16 value, qsizetype count)
 {

src/gui/painting/qdrawhelper_p.h

@@ -70,10 +70,17 @@ QT_BEGIN_NAMESPACE
 
 #if defined(Q_CC_GNU)
 #  define Q_DECL_RESTRICT __restrict__
+#  if defined(Q_PROCESSOR_X86_32) && defined(Q_CC_GNU) && !defined(Q_CC_CLANG) && !defined(Q_CC_INTEL)
+#    define Q_DECL_VECTORCALL __attribute__((sseregparm,regparm(3)))
+#  else
+#    define Q_DECL_VECTORCALL
+#  endif
 #elif defined(Q_CC_MSVC)
 #  define Q_DECL_RESTRICT __restrict
+#  define Q_DECL_VECTORCALL __vectorcall
 #else
 #  define Q_DECL_RESTRICT
+#  define Q_DECL_VECTORCALL
 #endif
 
 static const uint AMASK = 0xff000000;

src/gui/painting/qdrawhelper_sse2.cpp

@@ -233,19 +233,70 @@ void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, u
     }
 }
 
+static Q_NEVER_INLINE
+void Q_DECL_VECTORCALL qt_memfillXX_aligned(void *dest, __m128i value128, quintptr bytecount)
+{
+    __m128i *dst128 = reinterpret_cast<__m128i *>(dest);
+    __m128i *end128 = reinterpret_cast<__m128i *>(static_cast<uchar *>(dest) + bytecount);
+
+    while (dst128 + 4 <= end128) {
+        _mm_store_si128(dst128 + 0, value128);
+        _mm_store_si128(dst128 + 1, value128);
+        _mm_store_si128(dst128 + 2, value128);
+        _mm_store_si128(dst128 + 3, value128);
+        dst128 += 4;
+    }
+
+    bytecount %= 4 * sizeof(__m128i);
+    switch (bytecount / sizeof(__m128i)) {
+    case 3: _mm_store_si128(dst128++, value128); Q_FALLTHROUGH();
+    case 2: _mm_store_si128(dst128++, value128); Q_FALLTHROUGH();
+    case 1: _mm_store_si128(dst128++, value128);
+    }
+}
+
+void qt_memfill64(quint64 *dest, quint64 value, qsizetype count)
+{
+    quintptr misaligned = quintptr(dest) % sizeof(__m128i);
+    if (misaligned && count) {
+#if defined(Q_PROCESSOR_X86_32)
+        // Before SSE came out, the alignment of the stack used to be only 4
+        // bytes and some OS/ABIs (notably, code generated by MSVC) still only
+        // align to that. In any case, we cannot count on the alignment of
+        // quint64 to be 8 -- see QtPrivate::AlignOf_WorkaroundForI386Abi in
+        // qglobal.h.
+        //
+        // If the pointer is not aligned to at least 8 bytes, then we'll never
+        // in turn hit a multiple of 16 for the qt_memfillXX_aligned call
+        // below.
+        if (Q_UNLIKELY(misaligned % sizeof(quint64)))
+            return qt_memfill_template(dest, value, count);
+#endif
+
+        *dest++ = value;
+        --count;
+    }
+
+    if (count % 2) {
+        dest[count - 1] = value;
+        --count;
+    }
+
+    qt_memfillXX_aligned(dest, _mm_set1_epi64x(value), count * sizeof(quint64));
+}
+
 void qt_memfill32(quint32 *dest, quint32 value, qsizetype count)
 {
-    if (count < 7) {
+    if (count < 4) {
+        // this simplifies the code below: the first switch can fall through
+        // without checking the value of count
         switch (count) {
-        case 6: *dest++ = value; Q_FALLTHROUGH();
-        case 5: *dest++ = value; Q_FALLTHROUGH();
-        case 4: *dest++ = value; Q_FALLTHROUGH();
         case 3: *dest++ = value; Q_FALLTHROUGH();
         case 2: *dest++ = value; Q_FALLTHROUGH();
         case 1: *dest   = value;
         }
         return;
-    };
+    }
 
     const int align = (quintptr)(dest) & 0xf;
     switch (align) {
@@ -263,24 +314,7 @@ void qt_memfill32(quint32 *dest, quint32 value, qsizetype count)
         }
     }
 
-    int count128 = count / 4;
-    __m128i *dst128 = reinterpret_cast<__m128i*>(dest);
-    __m128i *end128 = dst128 + count128;
-    const __m128i value128 = _mm_set1_epi32(value);
-
-    while (dst128 + 3 < end128) {
-        _mm_store_si128(dst128 + 0, value128);
-        _mm_store_si128(dst128 + 1, value128);
-        _mm_store_si128(dst128 + 2, value128);
-        _mm_store_si128(dst128 + 3, value128);
-        dst128 += 4;
-    }
-
-    switch (count128 & 0x3) {
-    case 3:      _mm_store_si128(dst128++, value128); Q_FALLTHROUGH();
-    case 2:      _mm_store_si128(dst128++, value128); Q_FALLTHROUGH();
-    case 1:      _mm_store_si128(dst128++, value128);
-    }
+    qt_memfillXX_aligned(dest, _mm_set1_epi32(value), count * sizeof(quint32));
 }
 
 void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, uint color, uint const_alpha)

src/gui/painting/qdrawhelper_x86_p.h

@@ -57,6 +57,7 @@
 QT_BEGIN_NAMESPACE
 
 #ifdef __SSE2__
+void qt_memfill64(quint64 *dest, quint64 value, qsizetype count);
 void qt_memfill32(quint32 *dest, quint32 value, qsizetype count);
 void qt_bitmapblit32_sse2(QRasterBuffer *rasterBuffer, int x, int y,
                           const QRgba64 &color,