Avoid auto-vectorization of epilogues of manual vectorization

Defines a structure that tells the compiler in no uncertain terms the
maximum number of times a loop can be run.

The reduces the size of qdrawhelper_avx2.o from 22kbytes to 11kbytes.

Change-Id: Ie3d6281b04b4be3332497c15f3dfe9f185e20507
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>

8 files changed, 29 insertions, 33 deletions

diff --git a/src/corelib/tools/qsimd_p.h b/src/corelib/tools/qsimd_p.h
index 2fd4be00a5..3b6c85ca8a 100644
--- a/src/corelib/tools/qsimd_p.h
+++ b/src/corelib/tools/qsimd_p.h
@@ -471,6 +471,9 @@ static inline quint64 qCpuFeatures()
 #define ALIGNMENT_PROLOGUE_32BYTES(ptr, i, length) \
     for (; i < static_cast<int>(qMin(static_cast<quintptr>(length), ((8 - ((reinterpret_cast<quintptr>(ptr) >> 2) & 0x7)) & 0x7))); ++i)
 
+#define SIMD_EPILOGUE(i, length, max) \
+    for (int _i = 0; _i < max && i < length; ++i, ++_i)
+
 QT_END_NAMESPACE
 
 #endif // QSIMD_P_H
diff --git a/src/gui/image/qimage_ssse3.cpp b/src/gui/image/qimage_ssse3.cpp
index 0fa0eecd80..9cdfba20e3 100644
--- a/src/gui/image/qimage_ssse3.cpp
+++ b/src/gui/image/qimage_ssse3.cpp
@@ -50,15 +50,11 @@ QT_BEGIN_NAMESPACE
 // dst must be at least len * 4 bytes
 Q_GUI_EXPORT void QT_FASTCALL qt_convert_rgb888_to_rgb32_ssse3(quint32 *dst, const uchar *src, int len)
 {
-    quint32 *const end = dst + len;
+    int i = 0;
 
-    // Prologue, align dst to 16 bytes. The alignment is done on dst because it has 4 store()
-    // for each 3 load() of src.
-    const int offsetToAlignOn16Bytes = (4 - ((reinterpret_cast<quintptr>(dst) >> 2) & 0x3)) & 0x3;
-    const int prologLength = qMin(len, offsetToAlignOn16Bytes);
-
-    for (int i = 0; i < prologLength; ++i) {
-        *dst++ = qRgb(src[0], src[1], src[2]);
+    // Prologue, align dst to 16 bytes.
+    ALIGNMENT_PROLOGUE_16BYTES(dst, i, len) {
+        dst[i] = qRgb(src[0], src[1], src[2]);
         src += 3;
     }
 
@@ -72,10 +68,9 @@ Q_GUI_EXPORT void QT_FASTCALL qt_convert_rgb888_to_rgb32_ssse3(quint32 *dst, con
     const __m128i alphaMask = _mm_set1_epi32(0xff000000);
 
     const __m128i *inVectorPtr = (const __m128i *)src;
-    __m128i *dstVectorPtr = (__m128i *)dst;
+    __m128i *dstVectorPtr = (__m128i *)(dst + i);
 
-    const int simdRoundCount = (len - prologLength) / 16; // one iteration in the loop converts 16 pixels
-    for (int i = 0; i < simdRoundCount; ++i) {
+    for (; i < (len - 15); i += 16) { // one iteration in the loop converts 16 pixels
         /*
          RGB888 has 5 pixels per vector, + 1 byte from the next pixel. The idea here is
          to load vectors of RGB888 and use palignr to select a vector out of two vectors.
@@ -117,10 +112,9 @@ Q_GUI_EXPORT void QT_FASTCALL qt_convert_rgb888_to_rgb32_ssse3(quint32 *dst, con
         ++dstVectorPtr;
     }
     src = (const uchar *)inVectorPtr;
-    dst = (quint32 *)dstVectorPtr;
 
-    while (dst != end) {
-        *dst++ = qRgb(src[0], src[1], src[2]);
+    SIMD_EPILOGUE(i, len, 15) {
+        dst[i] = qRgb(src[0], src[1], src[2]);
         src += 3;
     }
 }
diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index 3e01d34cb2..928a56fd2f 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -603,7 +603,7 @@ static inline void qConvertARGB32PMToARGB64PM_sse2(QRgba64 *buffer, const uint *
         buffer += 2;
     }
 
-    for (; i < count; ++i) {
+    SIMD_EPILOGUE(i, count, 3) {
         uint s = *src++;
         if (RGBA)
             s = RGBA2ARGB(s);
@@ -766,7 +766,7 @@ static inline void qConvertA2RGB30PMToARGB64PM_sse2(QRgba64 *buffer, const uint
         buffer += 2;
     }
 
-    for (; i < count; ++i)
+    SIMD_EPILOGUE(i, count, 3)
         *buffer++ = qConvertA2rgb30ToRgb64<PixelOrder>(*src++);
 }
 #endif
@@ -1397,7 +1397,7 @@ static inline void qConvertARGB64PMToA2RGB30PM_sse2(uint *dest, const QRgba64 *b
         }
     }
 
-    for (; i < count; ++i)
+    SIMD_EPILOGUE(i, count, 15)
         *dest++ = qConvertRgb64ToRgb30<PixelOrder>(*buffer++);
 }
 #endif
diff --git a/src/gui/painting/qdrawhelper_avx2.cpp b/src/gui/painting/qdrawhelper_avx2.cpp
index b3fa380dc0..acc9bc7ba1 100644
--- a/src/gui/painting/qdrawhelper_avx2.cpp
+++ b/src/gui/painting/qdrawhelper_avx2.cpp
@@ -201,7 +201,7 @@ inline static void BLEND_SOURCE_OVER_ARGB32_WITH_CONST_ALPHA_AVX2(quint32 *dst,
             _mm256_store_si256((__m256i *)&dst[x], dstVector);
         }
     }
-    for (; x < length; ++x)
+    SIMD_EPILOGUE(x, length, 7)
         blend_pixel(dst[x], src[x], const_alpha);
 }
 
@@ -275,7 +275,7 @@ void qt_blend_rgb32_on_rgb32_avx2(uchar *destPixels, int dbpl,
         }
 
         // 3) Epilogue
-        for (; x < w; ++x)
+        SIMD_EPILOGUE(x, w, 7)
             dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
 
         srcPixels += sbpl;
@@ -322,7 +322,7 @@ void QT_FASTCALL comp_func_Source_avx2(uint *dst, const uint *src, int length, u
         }
 
         // 3) Epilogue
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 7)
             dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
     }
 }
@@ -353,7 +353,7 @@ void QT_FASTCALL comp_func_solid_SourceOver_avx2(uint *destPixels, int length, u
             dstVector = _mm256_add_epi8(colorVector, dstVector);
             _mm256_store_si256((__m256i *)&dst[x], dstVector);
         }
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 7)
             destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
     }
 }
diff --git a/src/gui/painting/qdrawhelper_neon.cpp b/src/gui/painting/qdrawhelper_neon.cpp
index f5b794ace6..d51b43961c 100644
--- a/src/gui/painting/qdrawhelper_neon.cpp
+++ b/src/gui/painting/qdrawhelper_neon.cpp
@@ -817,7 +817,7 @@ void QT_FASTCALL comp_func_solid_SourceOver_neon(uint *destPixels, int length, u
             vst1q_u32(&dst[x], colorPlusBlendedPixels);
         }
 
-        for (;x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
     }
 }
@@ -869,7 +869,7 @@ void QT_FASTCALL comp_func_Plus_neon(uint *dst, const uint *src, int length, uin
             vst1q_u32((uint32_t *)&dst[x], vcombine_u32(result32_low, result32_high));
         }
 
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
     }
 }
diff --git a/src/gui/painting/qdrawhelper_sse2.cpp b/src/gui/painting/qdrawhelper_sse2.cpp
index 03abeed440..5ff08e8153 100644
--- a/src/gui/painting/qdrawhelper_sse2.cpp
+++ b/src/gui/painting/qdrawhelper_sse2.cpp
@@ -126,9 +126,8 @@ void qt_blend_rgb32_on_rgb32_sse2(uchar *destPixels, int dbpl,
                         _mm_store_si128((__m128i *)&dst[x], result);
                     }
                 }
-                for (; x<w; ++x) {
+                SIMD_EPILOGUE(x, w, 3)
                     dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], one_minus_const_alpha);
-                }
                 dst = (quint32 *)(((uchar *) dst) + dbpl);
                 src = (const quint32 *)(((const uchar *) src) + sbpl);
             }
@@ -177,7 +176,7 @@ void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uin
         }
 
         // 3) Epilogue:
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             dst[x] = comp_func_Plus_one_pixel(dst[x], src[x]);
     } else {
         const int one_minus_const_alpha = 255 - const_alpha;
@@ -201,7 +200,7 @@ void QT_FASTCALL comp_func_Plus_sse2(uint *dst, const uint *src, int length, uin
         }
 
         // 3) Epilogue:
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             dst[x] = comp_func_Plus_one_pixel_const_alpha(dst[x], src[x], const_alpha, one_minus_const_alpha);
     }
 }
@@ -232,7 +231,7 @@ void QT_FASTCALL comp_func_Source_sse2(uint *dst, const uint *src, int length, u
         }
 
         // 3) Epilogue
-        for (; x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             dst[x] = INTERPOLATE_PIXEL_255(src[x], const_alpha, dst[x], ialpha);
     }
 }
@@ -313,7 +312,7 @@ void QT_FASTCALL comp_func_solid_SourceOver_sse2(uint *destPixels, int length, u
             dstVector = _mm_add_epi8(colorVector, dstVector);
             _mm_store_si128((__m128i *)&dst[x], dstVector);
         }
-        for (;x < length; ++x)
+        SIMD_EPILOGUE(x, length, 3)
             destPixels[x] = color + BYTE_MUL(destPixels[x], minusAlphaOfColor);
     }
 }
@@ -592,7 +591,7 @@ void qt_scale_image_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
             BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask);
         }
 
-        for (; x<w; x++) {
+        SIMD_EPILOGUE(x, w, 3) {
             uint s = src[(basex + x*ix) >> 16];
             dst[x] = s + BYTE_MUL(dst[x], qAlpha(~s));
         }
diff --git a/src/gui/painting/qdrawhelper_ssse3.cpp b/src/gui/painting/qdrawhelper_ssse3.cpp
index 2026a4e656..45ecc8b422 100644
--- a/src/gui/painting/qdrawhelper_ssse3.cpp
+++ b/src/gui/painting/qdrawhelper_ssse3.cpp
@@ -215,7 +215,7 @@ static inline void store_uint24_ssse3(uchar *dst, const uint *src, int len)
     dst24 = reinterpret_cast<quint24*>(dstVectorPtr);
     src = reinterpret_cast<const uint*>(inVectorPtr);
 
-    for (; i < len; ++i)
+    SIMD_EPILOGUE(i, len, 15)
         *dst24++ = quint24(*src++);
 }
 
diff --git a/src/gui/painting/qdrawingprimitive_sse2_p.h b/src/gui/painting/qdrawingprimitive_sse2_p.h
index 7affc63b32..93e4b9f572 100644
--- a/src/gui/painting/qdrawingprimitive_sse2_p.h
+++ b/src/gui/painting/qdrawingprimitive_sse2_p.h
@@ -178,7 +178,7 @@ QT_BEGIN_NAMESPACE
         const __m128i srcVector = _mm_loadu_si128((const __m128i *)&src[x]); \
         BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask) \
     } \
-    for (; x < length; ++x) { \
+    SIMD_EPILOGUE(x, length, 3) { \
         blend_pixel(dst[x], src[x]); \
     } \
 }
@@ -219,7 +219,7 @@ QT_BEGIN_NAMESPACE
             _mm_store_si128((__m128i *)&dst[x], result); \
         } \
     } \
-    for (; x < length; ++x) { \
+    SIMD_EPILOGUE(x, length, 3) { \
         blend_pixel(dst[x], src[x], const_alpha); \
     } \
 }