Split inner loop of Porter/Duff combiners out in an inline function rather than generating it in the macro

1 files changed, 80 insertions, 69 deletions

diff --git a/pixman/pixman-combine-float.c b/pixman/pixman-combine-float.c
index 484c77e..f58ce84 100644
--- a/pixman/pixman-combine-float.c
+++ b/pixman/pixman-combine-float.c
@@ -35,6 +35,84 @@
 
 #define IS_ZERO(f) ((f) == 0.0f) /* FIXME? */
 
+typedef float (* combine_channel_t) (float sa, float s, float da, float d);
+
+static force_inline void
+combine_inner (pixman_bool_t component,
+	       float *dest, const float *src, const float *mask, int n_pixels,
+	       combine_channel_t combine_a, combine_channel_t combine_c)
+{
+    int i;
+
+    if (!mask)
+    {
+	for (i = 0; i < 4 * n_pixels; i += 4)
+	{
+	    float sa = src[i + 0];
+	    float sr = src[i + 1];
+	    float sg = src[i + 2];
+	    float sb = src[i + 3];
+	    
+	    float da = dest[i + 0];
+	    float dr = dest[i + 1];
+	    float dg = dest[i + 2];
+	    float db = dest[i + 3];					
+	    
+	    dest[i + 0] = combine_a (sa, sa, da, da);
+	    dest[i + 1] = combine_c (sa, sr, da, dr);
+	    dest[i + 2] = combine_c (sa, sg, da, dg);
+	    dest[i + 3] = combine_c (sa, sb, da, db);
+	}
+    }
+    else
+    {
+	for (i = 0; i < 4 * n_pixels; i += 4)
+	{
+	    float sa, sr, sg, sb;
+	    float ma, mr, mg, mb;
+	    float da, dr, dg, db;
+	    
+	    sa = src[i + 0];
+	    sr = src[i + 1];
+	    sg = src[i + 2];
+	    sb = src[i + 3];
+	    
+	    if (component)
+	    {
+		ma = mask[i + 0];
+		mr = mask[i + 1];
+		mg = mask[i + 2];
+		mb = mask[i + 3];
+
+		sr *= mr;
+		sg *= mg;
+		sb *= mb;
+
+		ma *= sa;
+		mr *= sa;
+		mg *= sa;
+		mb *= sa;
+		
+		sa = ma;
+	    }
+	    else
+	    {
+		ma = mr = mg = mb = sa = mask[i + 0] * sa;
+	    }
+	    
+	    da = dest[i + 0];
+	    dr = dest[i + 1];
+	    dg = dest[i + 2];
+	    db = dest[i + 3];
+	    
+	    dest[i + 0] = combine_a (ma, sa, da, da);
+	    dest[i + 1] = combine_c (mr, sr, da, dr);
+	    dest[i + 2] = combine_c (mg, sg, da, dg);
+	    dest[i + 3] = combine_c (mb, sb, da, db);
+	}
+    }
+}
+
 #define MAKE_COMBINER(name, component, combine_a, combine_c)		\
     static void								\
     combine_ ## name ## _float (pixman_implementation_t *imp,		\
@@ -44,75 +122,8 @@
 				const float             *mask,		\
 				int		         n_pixels)	\
     {									\
-	int i;								\
-									\
-	if (!mask)							\
-	{								\
-	    for (i = 0; i < 4 * n_pixels; i += 4)			\
-	    {								\
-		float sa = src[i + 0];					\
-		float sr = src[i + 1];					\
-		float sg = src[i + 2];					\
-		float sb = src[i + 3];					\
-									\
-		float da = dest[i + 0];					\
-		float dr = dest[i + 1];					\
-		float dg = dest[i + 2];					\
-		float db = dest[i + 3];					\
-									\
-		dest[i + 0] = combine_a (sa, sa, da, da);		\
-		dest[i + 1] = combine_c (sa, sr, da, dr);		\
-		dest[i + 2] = combine_c (sa, sg, da, dg);		\
-		dest[i + 3] = combine_c (sa, sb, da, db);		\
-	    }								\
-	}								\
-	else								\
-	{								\
-	    for (i = 0; i < 4 * n_pixels; i += 4)			\
-	    {								\
-		float sa, sr, sg, sb;					\
-		float ma, mr, mg, mb;					\
-		float da, dr, dg, db;					\
-									\
-		sa = src[i + 0];					\
-		sr = src[i + 1];					\
-		sg = src[i + 2];					\
-		sb = src[i + 3];					\
-									\
-		if (component)						\
-		{							\
-		    ma = mask[i + 0];					\
-		    mr = mask[i + 1];					\
-		    mg = mask[i + 2];					\
-		    mb = mask[i + 3];					\
-		    							\
-		    sr *= mr;						\
-		    sg *= mg;						\
-		    sb *= mb;						\
-		    							\
-		    ma *= sa;						\
-		    mr *= sa;						\
-		    mg *= sa;						\
-		    mb *= sa;						\
-		    							\
-		    sa = ma;						\
-		}							\
-		else							\
-		{							\
-		    ma = mr = mg = mb = sa = mask[i + 0] * sa;		\
-		}							\
-									\
-		da = dest[i + 0];					\
-		dr = dest[i + 1];					\
-		dg = dest[i + 2];					\
-		db = dest[i + 3];					\
-									\
-		dest[i + 0] = combine_a (ma, sa, da, da);		\
-		dest[i + 1] = combine_c (mr, sr, da, dr);		\
-		dest[i + 2] = combine_c (mg, sg, da, dg);		\
-		dest[i + 3] = combine_c (mb, sb, da, db);		\
-	    }								\
-	}								\
+	combine_inner (component, dest, src, mask, n_pixels,		\
+		       combine_a, combine_c);				\
     }
 
 #define MAKE_COMBINERS(name, combine_a, combine_c)			\