Replace the inline blend function by #define

Some compilers do not inline the functions, which is a problem
because the number of arguments exceed the limit for SSE,
and because it is a lot slower for those low level functions.

1 files changed, 60 insertions, 67 deletions

diff --git a/src/gui/painting/qdrawhelper_sse2.cpp b/src/gui/painting/qdrawhelper_sse2.cpp
index 1dba914..6ac64d3 100644
--- a/src/gui/painting/qdrawhelper_sse2.cpp
+++ b/src/gui/painting/qdrawhelper_sse2.cpp
@@ -63,36 +63,36 @@ QT_BEGIN_NAMESPACE
  * colorMask must have 0x00ff00ff on each 32 bits component
  * half must have the value 128 (0x80) for each 32 bits compnent
  */
-Q_STATIC_INLINE_FUNCTION __m128i BYTE_MUL_SSE2(const __m128i pixelVector, const __m128i alphaChannel, const __m128i colorMask, const __m128i half)
-{
-    // 1. separate the colors in 2 vectors so each color is on 16 bits
-    // (in order to be multiplied by the alpha
-    // each 32 bit of dstVectorAG are in the form 0x00AA00GG
-    // each 32 bit of dstVectorRB are in the form 0x00RR00BB
-    __m128i pixelVectorAG = _mm_srli_epi16(pixelVector, 8);
-    __m128i pixelVectorRB = _mm_and_si128(pixelVector, colorMask);
-
-    // 2. multiply the vectors by the alpha channel
-    pixelVectorAG = _mm_mullo_epi16(pixelVectorAG, alphaChannel);
-    pixelVectorRB = _mm_mullo_epi16(pixelVectorRB, alphaChannel);
-
-    // 3. devide by 255, that's the tricky part.
-    // we do it like for BYTE_MUL(), with bit shift: X/255 ~= (X + X/256 + rounding)/256
-    /// so first (X + X/256 + rounding)
-    pixelVectorRB = _mm_add_epi16(pixelVectorRB, _mm_srli_epi16(pixelVectorRB, 8));
-    pixelVectorRB = _mm_add_epi16(pixelVectorRB, half);
-    pixelVectorAG = _mm_add_epi16(pixelVectorAG, _mm_srli_epi16(pixelVectorAG, 8));
-    pixelVectorAG = _mm_add_epi16(pixelVectorAG, half);
-
-    /// second devide by 256
-    pixelVectorRB = _mm_srli_epi16(pixelVectorRB, 8);
-    /// for AG, we could >> 8 to divide followed by << 8 to put the
-    /// bytes in the correct position. By masking instead, we execute
-    /// only one instruction
-    pixelVectorAG = _mm_andnot_si128(colorMask, pixelVectorAG);
-
-    // 4. combine the 2 pairs of colors
-    return _mm_or_si128(pixelVectorAG, pixelVectorRB);
+#define BYTE_MUL_SSE2(result, pixelVector, alphaChannel, colorMask, half) \
+{ \
+    /* 1. separate the colors in 2 vectors so each color is on 16 bits \
+       (in order to be multiplied by the alpha \
+       each 32 bit of dstVectorAG are in the form 0x00AA00GG \
+       each 32 bit of dstVectorRB are in the form 0x00RR00BB */\
+    __m128i pixelVectorAG = _mm_srli_epi16(pixelVector, 8); \
+    __m128i pixelVectorRB = _mm_and_si128(pixelVector, colorMask); \
+ \
+    /* 2. multiply the vectors by the alpha channel */\
+    pixelVectorAG = _mm_mullo_epi16(pixelVectorAG, alphaChannel); \
+    pixelVectorRB = _mm_mullo_epi16(pixelVectorRB, alphaChannel); \
+ \
+    /* 3. devide by 255, that's the tricky part. \
+       we do it like for BYTE_MUL(), with bit shift: X/255 ~= (X + X/256 + rounding)/256 */ \
+    /** so first (X + X/256 + rounding) */\
+    pixelVectorRB = _mm_add_epi16(pixelVectorRB, _mm_srli_epi16(pixelVectorRB, 8)); \
+    pixelVectorRB = _mm_add_epi16(pixelVectorRB, half); \
+    pixelVectorAG = _mm_add_epi16(pixelVectorAG, _mm_srli_epi16(pixelVectorAG, 8)); \
+    pixelVectorAG = _mm_add_epi16(pixelVectorAG, half); \
+ \
+    /** second devide by 256 */\
+    pixelVectorRB = _mm_srli_epi16(pixelVectorRB, 8); \
+    /** for AG, we could >> 8 to divide followed by << 8 to put the \
+        bytes in the correct position. By masking instead, we execute \
+        only one instruction */\
+    pixelVectorAG = _mm_andnot_si128(colorMask, pixelVectorAG); \
+ \
+    /* 4. combine the 2 pairs of colors */ \
+    result = _mm_or_si128(pixelVectorAG, pixelVectorRB); \
 }
 
 /*
@@ -101,34 +101,29 @@ Q_STATIC_INLINE_FUNCTION __m128i BYTE_MUL_SSE2(const __m128i pixelVector, const
  * colorMask must have 0x00ff00ff on each 32 bits component
  * half must have the value 128 (0x80) for each 32 bits compnent
  */
-Q_STATIC_INLINE_FUNCTION __m128i INTERPOLATE_PIXEL_255_SSE2(const __m128i srcVector,
-                                                            const __m128i dstVector,
-                                                            const __m128i alphaChannel,
-                                                            const __m128i oneMinusAlphaChannel ,
-                                                            const __m128i colorMask,
-                                                            const __m128i half) {
-    // interpolate AG
-    __m128i srcVectorAG = _mm_srli_epi16(srcVector, 8);
-    __m128i dstVectorAG = _mm_srli_epi16(dstVector, 8);
-    __m128i srcVectorAGalpha = _mm_mullo_epi16(srcVectorAG, alphaChannel);
-    __m128i dstVectorAGoneMinusAlphalpha = _mm_mullo_epi16(dstVectorAG, oneMinusAlphaChannel);
-    __m128i finalAG = _mm_add_epi16(srcVectorAGalpha, dstVectorAGoneMinusAlphalpha);
-    finalAG = _mm_add_epi16(finalAG, _mm_srli_epi16(finalAG, 8));
-    finalAG = _mm_add_epi16(finalAG, half);
-    finalAG = _mm_andnot_si128(colorMask, finalAG);
-
-    // interpolate RB
-    __m128i srcVectorRB = _mm_and_si128(srcVector, colorMask);
-    __m128i dstVectorRB = _mm_and_si128(dstVector, colorMask);
-    __m128i srcVectorRBalpha = _mm_mullo_epi16(srcVectorRB, alphaChannel);
-    __m128i dstVectorRBoneMinusAlphalpha = _mm_mullo_epi16(dstVectorRB, oneMinusAlphaChannel);
-    __m128i finalRB = _mm_add_epi16(srcVectorRBalpha, dstVectorRBoneMinusAlphalpha);
-    finalRB = _mm_add_epi16(finalRB, _mm_srli_epi16(finalRB, 8));
-    finalRB = _mm_add_epi16(finalRB, half);
-    finalRB = _mm_srli_epi16(finalRB, 8);
-
-    // combine
-    return _mm_or_si128(finalAG, finalRB);
+#define INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, alphaChannel, oneMinusAlphaChannel, colorMask, half) { \
+    /* interpolate AG */\
+    __m128i srcVectorAG = _mm_srli_epi16(srcVector, 8); \
+    __m128i dstVectorAG = _mm_srli_epi16(dstVector, 8); \
+    __m128i srcVectorAGalpha = _mm_mullo_epi16(srcVectorAG, alphaChannel); \
+    __m128i dstVectorAGoneMinusAlphalpha = _mm_mullo_epi16(dstVectorAG, oneMinusAlphaChannel); \
+    __m128i finalAG = _mm_add_epi16(srcVectorAGalpha, dstVectorAGoneMinusAlphalpha); \
+    finalAG = _mm_add_epi16(finalAG, _mm_srli_epi16(finalAG, 8)); \
+    finalAG = _mm_add_epi16(finalAG, half); \
+    finalAG = _mm_andnot_si128(colorMask, finalAG); \
+ \
+    /* interpolate RB */\
+    __m128i srcVectorRB = _mm_and_si128(srcVector, colorMask); \
+    __m128i dstVectorRB = _mm_and_si128(dstVector, colorMask); \
+    __m128i srcVectorRBalpha = _mm_mullo_epi16(srcVectorRB, alphaChannel); \
+    __m128i dstVectorRBoneMinusAlphalpha = _mm_mullo_epi16(dstVectorRB, oneMinusAlphaChannel); \
+    __m128i finalRB = _mm_add_epi16(srcVectorRBalpha, dstVectorRBoneMinusAlphalpha); \
+    finalRB = _mm_add_epi16(finalRB, _mm_srli_epi16(finalRB, 8)); \
+    finalRB = _mm_add_epi16(finalRB, half); \
+    finalRB = _mm_srli_epi16(finalRB, 8); \
+ \
+    /* combine */\
+    result = _mm_or_si128(finalAG, finalRB); \
 }
 
 void qt_blend_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
@@ -165,7 +160,8 @@ void qt_blend_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
                     alphaChannel = _mm_sub_epi16(one, alphaChannel);
 
                     const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
-                    const __m128i destMultipliedByOneMinusAlpha = BYTE_MUL_SSE2(dstVector, alphaChannel, colorMask, half);
+                    __m128i destMultipliedByOneMinusAlpha;
+                    BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half);
 
                     // result = s + d * (1-alpha)
                     const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha);
@@ -197,14 +193,15 @@ void qt_blend_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
             for (; x < w-3; x += 4) {
                 __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
                 if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) {
-                    srcVector = BYTE_MUL_SSE2(srcVector, constAlphaVector, colorMask, half);
+                    BYTE_MUL_SSE2(srcVector, srcVector, constAlphaVector, colorMask, half);
 
                     __m128i alphaChannel = _mm_srli_epi32(srcVector, 24);
                     alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16));
                     alphaChannel = _mm_sub_epi16(one, alphaChannel);
 
                     const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
-                    const __m128i destMultipliedByOneMinusAlpha = BYTE_MUL_SSE2(dstVector, alphaChannel, colorMask, half);
+                    __m128i destMultipliedByOneMinusAlpha;
+                    BYTE_MUL_SSE2(destMultipliedByOneMinusAlpha, dstVector, alphaChannel, colorMask, half);
 
                     const __m128i result = _mm_add_epi8(srcVector, destMultipliedByOneMinusAlpha);
                     _mm_storeu_si128((__m128i *)&dst[x], result);
@@ -252,12 +249,8 @@ void qt_blend_rgb32_on_rgb32_sse2(uchar *destPixels, int dbpl,
                     __m128i srcVector = _mm_loadu_si128((__m128i *)&src[x]);
                     if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVector, nullVector)) != 0xffff) {
                         const __m128i dstVector = _mm_loadu_si128((__m128i *)&dst[x]);
-                        const __m128i result = INTERPOLATE_PIXEL_255_SSE2(srcVector,
-                                                                          dstVector,
-                                                                          constAlphaVector,
-                                                                          oneMinusConstAlpha,
-                                                                          colorMask,
-                                                                          half);
+                        __m128i result;
+                        INTERPOLATE_PIXEL_255_SSE2(result, srcVector, dstVector, constAlphaVector, oneMinusConstAlpha, colorMask, half);
                         _mm_storeu_si128((__m128i *)&dst[x], result);
                     }
                 }