changeset:   96613:3e003f0b8026
tag:         2pass
tag:         qbase
tag:         qtip
tag:         tip
user:        Jeff Muizelaar <jmuizelaar@mozilla.com>
date:        Thu May 17 19:23:53 2012 -0400
summary:     Bug 757878. Add a fast path for 8888_over_565 with NEON. r=bgirard,joe

diff --git a/gfx/cairo/libpixman/src/pixman-arm-common.h b/gfx/cairo/libpixman/src/pixman-arm-common.h
--- a/gfx/cairo/libpixman/src/pixman-arm-common.h
+++ b/gfx/cairo/libpixman/src/pixman-arm-common.h
@@ -355,26 +355,26 @@ scaled_bilinear_scanline_##cputype##_##n
     if ((flags & SKIP_ZERO_SRC) && zero_src)                                  \
 	return;                                                               \
     pixman_scaled_bilinear_scanline_##name##_##op##_asm_##cputype (           \
                             dst, src_top, src_bottom, wt, wb, vx, unit_x, w); \
 }                                                                             \
                                                                               \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_cover_##op,                 \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint32_t, dst_type, COVER, FLAG_NONE)        \
+                       NULL, src_type, uint32_t, dst_type, COVER, FLAG_NONE)  \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_none_##op,                  \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint32_t, dst_type, NONE, FLAG_NONE)         \
+                       NULL, src_type, uint32_t, dst_type, NONE, FLAG_NONE)   \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_pad_##op,                   \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint32_t, dst_type, PAD, FLAG_NONE)          \
+                       NULL, src_type, uint32_t, dst_type, PAD, FLAG_NONE)    \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_normal_##op,                \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint32_t, dst_type, NORMAL,                  \
+                       NULL, src_type, uint32_t, dst_type, NORMAL,            \
                        FLAG_NONE)
 
 
 #define PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_A8_DST(flags, cputype, name, op,  \
                                                 src_type, dst_type)           \
 void                                                                          \
 pixman_scaled_bilinear_scanline_##name##_##op##_asm_##cputype (               \
                                                 dst_type *       dst,         \
@@ -404,25 +404,25 @@ scaled_bilinear_scanline_##cputype##_##n
     if ((flags & SKIP_ZERO_SRC) && zero_src)                                  \
 	return;                                                                   \
     pixman_scaled_bilinear_scanline_##name##_##op##_asm_##cputype (           \
                       dst, mask, src_top, src_bottom, wt, wb, vx, unit_x, w); \
 }                                                                             \
                                                                               \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_cover_##op,                 \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint8_t, dst_type, COVER,                    \
+                       NULL, src_type, uint8_t, dst_type, COVER,              \
                        FLAG_HAVE_NON_SOLID_MASK)                              \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_none_##op,                  \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint8_t, dst_type, NONE,                     \
+                       NULL, src_type, uint8_t, dst_type, NONE,               \
                        FLAG_HAVE_NON_SOLID_MASK)                              \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_pad_##op,                   \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint8_t, dst_type, PAD,                      \
+                       NULL, src_type, uint8_t, dst_type, PAD,                \
                        FLAG_HAVE_NON_SOLID_MASK)                              \
 FAST_BILINEAR_MAINLOOP_COMMON (cputype##_##name##_normal_##op,                \
                        scaled_bilinear_scanline_##cputype##_##name##_##op,    \
-                       src_type, uint8_t, dst_type, NORMAL,                   \
+                       NULL, src_type, uint8_t, dst_type, NORMAL,             \
                        FLAG_HAVE_NON_SOLID_MASK)
 
 
 #endif
diff --git a/gfx/cairo/libpixman/src/pixman-arm-neon.c b/gfx/cairo/libpixman/src/pixman-arm-neon.c
--- a/gfx/cairo/libpixman/src/pixman-arm-neon.c
+++ b/gfx/cairo/libpixman/src/pixman-arm-neon.c
@@ -140,16 +140,33 @@ PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST 
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 8888_0565, SRC,
                                          uint32_t, uint16_t)
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 0565_x888, SRC,
                                          uint16_t, uint32_t)
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (0, neon, 0565_0565, SRC,
                                          uint16_t, uint16_t)
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (SKIP_ZERO_SRC, neon, 8888_8888, OVER,
                                          uint32_t, uint32_t)
+static force_inline void
+pixman_scaled_bilinear_scanline_8888_8888_SRC (
+                                                uint32_t *       dst,
+                                                const uint32_t * mask,
+                                                const uint32_t * src_top,
+                                                const uint32_t * src_bottom,
+                                                int32_t          w,
+                                                int              wt,
+                                                int              wb,
+                                                pixman_fixed_t   vx,
+                                                pixman_fixed_t   unit_x,
+                                                pixman_fixed_t   max_vx,
+                                                pixman_bool_t    zero_src)
+{
+    pixman_scaled_bilinear_scanline_8888_8888_SRC_asm_neon (dst, src_top, src_bottom, wt, wb, vx, unit_x, w);
+}
+
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_DST (SKIP_ZERO_SRC, neon, 8888_8888, ADD,
                                          uint32_t, uint32_t)
 
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_A8_DST (0, neon, 8888_8_8888, SRC,
                                             uint32_t, uint32_t)
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_A8_DST (0, neon, 8888_8_0565, SRC,
                                             uint32_t, uint16_t)
 PIXMAN_ARM_BIND_SCALED_BILINEAR_SRC_A8_DST (0, neon, 0565_8_x888, SRC,
@@ -261,16 +278,38 @@ pixman_blt_neon (uint32_t *src_bits,
 		(uint32_t *)(((char *) src_bits) +
 		src_y * src_stride * 4 + src_x * 4), src_stride);
 	return TRUE;
     default:
 	return FALSE;
     }
 }
 
+static inline void op_bilinear_over_8888_0565(uint16_t *dst, const uint32_t *mask, const uint32_t *src, int width)
+{
+    pixman_composite_over_8888_0565_asm_neon (width, 1, dst, 0, src, 0);
+}
+
+FAST_BILINEAR_MAINLOOP_COMMON (neon_8888_0565_cover_OVER,
+			       pixman_scaled_bilinear_scanline_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       COVER, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (neon_8888_0565_pad_OVER,
+			       pixman_scaled_bilinear_scanline_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       PAD, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (neon_8888_0565_none_OVER,
+			       pixman_scaled_bilinear_scanline_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       NONE, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (neon_8888_0565_normal_OVER,
+			       pixman_scaled_bilinear_scanline_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       NORMAL, FLAG_NONE)
+
 static const pixman_fast_path_t arm_neon_fast_paths[] =
 {
     PIXMAN_STD_FAST_PATH (SRC,  r5g6b5,   null,     r5g6b5,   neon_composite_src_0565_0565),
     PIXMAN_STD_FAST_PATH (SRC,  b5g6r5,   null,     b5g6r5,   neon_composite_src_0565_0565),
     PIXMAN_STD_FAST_PATH (SRC,  a8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
     PIXMAN_STD_FAST_PATH (SRC,  x8r8g8b8, null,     r5g6b5,   neon_composite_src_8888_0565),
     PIXMAN_STD_FAST_PATH (SRC,  a8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
     PIXMAN_STD_FAST_PATH (SRC,  x8b8g8r8, null,     b5g6r5,   neon_composite_src_8888_0565),
@@ -414,16 +453,18 @@ static const pixman_fast_path_t arm_neon
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (SRC, r5g6b5, r5g6b5, neon_0565_8_0565),
 
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, neon_8888_8_8888),
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8r8g8b8, x8r8g8b8, neon_8888_8_8888),
 
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (ADD, a8r8g8b8, a8r8g8b8, neon_8888_8_8888),
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (ADD, a8r8g8b8, x8r8g8b8, neon_8888_8_8888),
 
+    SIMPLE_BILINEAR_FAST_PATH (OVER, a8r8g8b8, r5g6b5, neon_8888_0565),
+
     { PIXMAN_OP_NONE },
 };
 
 static pixman_bool_t
 arm_neon_blt (pixman_implementation_t *imp,
               uint32_t *               src_bits,
               uint32_t *               dst_bits,
               int                      src_stride,
diff --git a/gfx/cairo/libpixman/src/pixman-fast-path.c b/gfx/cairo/libpixman/src/pixman-fast-path.c
--- a/gfx/cairo/libpixman/src/pixman-fast-path.c
+++ b/gfx/cairo/libpixman/src/pixman-fast-path.c
@@ -1356,63 +1356,63 @@ scaled_bilinear_scanline_565_565_SRC (ui
         vx += unit_x;
         *dst++ = d;
     }
 }
 
 #endif
 
 FAST_BILINEAR_MAINLOOP_COMMON (565_565_cover_SRC,
-			       scaled_bilinear_scanline_565_565_SRC,
+			       scaled_bilinear_scanline_565_565_SRC, NULL,
 			       uint16_t, uint32_t, uint16_t,
 			       COVER, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (565_565_pad_SRC,
-			       scaled_bilinear_scanline_565_565_SRC,
+			       scaled_bilinear_scanline_565_565_SRC, NULL,
 			       uint16_t, uint32_t, uint16_t,
 			       PAD, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (565_565_none_SRC,
-			       scaled_bilinear_scanline_565_565_SRC,
+			       scaled_bilinear_scanline_565_565_SRC, NULL,
 			       uint16_t, uint32_t, uint16_t,
 			       NONE, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (565_565_normal_SRC,
-			       scaled_bilinear_scanline_565_565_SRC,
+			       scaled_bilinear_scanline_565_565_SRC, NULL,
 			       uint16_t, uint32_t, uint16_t,
 			       NORMAL, FLAG_NONE)
 
 FAST_BILINEAR_MAINLOOP_COMMON (8888_565_cover_OVER,
-			       scaled_bilinear_scanline_8888_565_OVER,
+			       scaled_bilinear_scanline_8888_565_OVER, NULL,
 			       uint32_t, uint32_t, uint16_t,
 			       COVER, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_565_pad_OVER,
-			       scaled_bilinear_scanline_8888_565_OVER,
+			       scaled_bilinear_scanline_8888_565_OVER, NULL,
 			       uint32_t, uint32_t, uint16_t,
 			       PAD, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_565_none_OVER,
-			       scaled_bilinear_scanline_8888_565_OVER,
+			       scaled_bilinear_scanline_8888_565_OVER, NULL,
 			       uint32_t, uint32_t, uint16_t,
 			       NONE, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_565_normal_OVER,
-			       scaled_bilinear_scanline_8888_565_OVER,
+			       scaled_bilinear_scanline_8888_565_OVER, NULL,
 			       uint32_t, uint32_t, uint16_t,
 			       NORMAL, FLAG_NONE)
 
 FAST_BILINEAR_MAINLOOP_COMMON (8888_8888_cover_OVER,
-			       scaled_bilinear_scanline_8888_8888_OVER,
+			       scaled_bilinear_scanline_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       COVER, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_8888_pad_OVER,
-			       scaled_bilinear_scanline_8888_8888_OVER,
+			       scaled_bilinear_scanline_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       PAD, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_8888_none_OVER,
-			       scaled_bilinear_scanline_8888_8888_OVER,
+			       scaled_bilinear_scanline_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NONE, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (8888_8888_normal_OVER,
-			       scaled_bilinear_scanline_8888_8888_OVER,
+			       scaled_bilinear_scanline_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NORMAL, FLAG_NONE)
 
 #define REPEAT_MIN_WIDTH    32
 
 static void
 fast_composite_tiled_repeat (pixman_implementation_t *imp,
 			     pixman_composite_info_t *info)
diff --git a/gfx/cairo/libpixman/src/pixman-inlines.h b/gfx/cairo/libpixman/src/pixman-inlines.h
--- a/gfx/cairo/libpixman/src/pixman-inlines.h
+++ b/gfx/cairo/libpixman/src/pixman-inlines.h
@@ -816,18 +816,48 @@ bilinear_pad_repeat_get_scanline_bounds 
  *
  * Note: normally the sum of 'weight_top' and 'weight_bottom' is equal to 256,
  *       but sometimes it may be less than that for NONE repeat when handling
  *       fuzzy antialiased top or bottom image edges. Also both top and
  *       bottom weight variables are guaranteed to have value in 0-255
  *       range and can fit into unsigned byte or be used with 8-bit SIMD
  *       multiplication instructions.
  */
-#define FAST_BILINEAR_MAINLOOP_INT(scale_func_name, scanline_func, src_type_t, mask_type_t,	\
-				  dst_type_t, repeat_mode, flags)				\
+
+/* Replace a single "scanline_func" with "fetch_func" & "op_func" to allow optional
+ * two stage processing (bilinear fetch to a temp buffer, followed by unscaled
+ * combine), "op_func" may be NULL, in this case we keep old behavior.
+ * This is ugly and gcc issues some warnings, but works.
+ *
+ * An advice: clang has much better error reporting than gcc for deeply nested macros.
+ */
+
+#define	scanline_func(dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,            \
+                      scanline_buf, mask, src_top, src_bottom, width,                           \
+                      weight_top, weight_bottom, vx, unit_x, max_vx, zero_src)                  \
+ do {                                                                                           \
+		if (op_func != NULL)								\
+		{										\
+		    fetch_func ((void *)scanline_buf, (mask), (src_top), (src_bottom), (width), \
+                        (weight_top), (weight_bottom), (vx), (unit_x), (max_vx), (zero_src));   \
+		    ((void (*)(dst_type_t *, const mask_type_t *, const src_type_t *, int)) op_func)\
+			((dst), (mask), (src_type_t *)scanline_buf, (width));			\
+		}										\
+		else										\
+		{										\
+		    fetch_func ((void*)(dst), (mask), (src_top), (src_bottom), (width), (weight_top),  \
+                                (weight_bottom), (vx), (unit_x), (max_vx), (zero_src));         \
+		}                                                                               \
+  } while (0)
+
+
+#define SCANLINE_BUFFER_LENGTH 3072
+
+#define FAST_BILINEAR_MAINLOOP_INT(scale_func_name, fetch_func, op_func, src_type_t,		\
+				  mask_type_t, dst_type_t, repeat_mode, flags)			\
 static void											\
 fast_composite_scaled_bilinear ## scale_func_name (pixman_implementation_t *imp,		\
 						   pixman_composite_info_t *info)		\
 {												\
     PIXMAN_COMPOSITE_ARGS (info);								\
     dst_type_t *dst_line;									\
     mask_type_t *mask_line;									\
     src_type_t *src_first_line;									\
@@ -842,16 +872,19 @@ fast_composite_scaled_bilinear ## scale_
     mask_type_t solid_mask;									\
     const mask_type_t *mask = &solid_mask;							\
     int src_stride, mask_stride, dst_stride;							\
 												\
     int src_width;										\
     pixman_fixed_t src_width_fixed;								\
     int max_x;											\
     pixman_bool_t need_src_extension;								\
+                                                                                                \
+    uint64_t stack_scanline_buffer[SCANLINE_BUFFER_LENGTH];                                     \
+    uint8_t *scanline_buffer = (uint8_t *) stack_scanline_buffer;                               \
 												\
     PIXMAN_IMAGE_GET_LINE (dest_image, dest_x, dest_y, dst_type_t, dst_stride, dst_line, 1);	\
     if (flags & FLAG_HAVE_SOLID_MASK)								\
     {												\
 	solid_mask = _pixman_image_get_solid (imp, mask_image, dest_image->bits.format);	\
 	mask_stride = 0;									\
     }												\
     else if (flags & FLAG_HAVE_NON_SOLID_MASK)							\
@@ -914,16 +947,24 @@ fast_composite_scaled_bilinear ## scale_
 	else											\
 	{											\
 	    src_width = src_image->bits.width;							\
 	    need_src_extension = FALSE;								\
 	}											\
 												\
 	src_width_fixed = pixman_int_to_fixed (src_width);					\
     }												\
+                                                                                                \
+    if (op_func != NULL && width * sizeof(src_type_t) > sizeof(stack_scanline_buffer))          \
+    {                                                                                           \
+	scanline_buffer = pixman_malloc_ab (width, sizeof(src_type_t));                         \
+                                                                                                \
+	if (!scanline_buffer)                                                                   \
+	    return;                                                                             \
+    }                                                                                           \
 												\
     while (--height >= 0)									\
     {												\
 	int weight1, weight2;									\
 	dst = dst_line;										\
 	dst_line += dst_stride;									\
 	vx = v.vector[0];									\
 	if (flags & FLAG_HAVE_NON_SOLID_MASK)							\
@@ -956,36 +997,39 @@ fast_composite_scaled_bilinear ## scale_
 	    repeat (PIXMAN_REPEAT_PAD, &y2, src_image->bits.height);				\
 	    src1 = src_first_line + src_stride * y1;						\
 	    src2 = src_first_line + src_stride * y2;						\
 												\
 	    if (left_pad > 0)									\
 	    {											\
 		buf1[0] = buf1[1] = src1[0];							\
 		buf2[0] = buf2[1] = src2[0];							\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, left_pad, weight1, weight2, 0, 0, 0, FALSE);		\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, left_pad, weight1, weight2,   \
+                               0, 0, 0, FALSE);	                                                \
 		dst += left_pad;								\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += left_pad;								\
 	    }											\
 	    if (width > 0)									\
 	    {											\
-		scanline_func (dst, mask,							\
-			       src1, src2, width, weight1, weight2, vx, unit_x, 0, FALSE);	\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, src1, src2, width, weight1, weight2,      \
+                               vx, unit_x, 0, FALSE);                                           \
 		dst += width;									\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += width;								\
 	    }											\
 	    if (right_pad > 0)									\
 	    {											\
 		buf1[0] = buf1[1] = src1[src_image->bits.width - 1];				\
 		buf2[0] = buf2[1] = src2[src_image->bits.width - 1];				\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, right_pad, weight1, weight2, 0, 0, 0, FALSE);	\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, right_pad, weight1, weight2,  \
+                               0, 0, 0, FALSE);                                                 \
 	    }											\
 	}											\
 	else if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NONE)				\
 	{											\
 	    src_type_t *src1, *src2;								\
 	    src_type_t buf1[2];									\
 	    src_type_t buf2[2];									\
 	    /* handle top/bottom zero padding by just setting weights to 0 if needed */		\
@@ -1011,64 +1055,67 @@ fast_composite_scaled_bilinear ## scale_
 	    }											\
 	    src1 = src_first_line + src_stride * y1;						\
 	    src2 = src_first_line + src_stride * y2;						\
 												\
 	    if (left_pad > 0)									\
 	    {											\
 		buf1[0] = buf1[1] = 0;								\
 		buf2[0] = buf2[1] = 0;								\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, left_pad, weight1, weight2, 0, 0, 0, TRUE);		\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, left_pad, weight1, weight2,   \
+                               0, 0, 0, TRUE);	                                                \
 		dst += left_pad;								\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += left_pad;								\
 	    }											\
 	    if (left_tz > 0)									\
 	    {											\
 		buf1[0] = 0;									\
 		buf1[1] = src1[0];								\
 		buf2[0] = 0;									\
 		buf2[1] = src2[0];								\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, left_tz, weight1, weight2,				\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, left_tz, weight1, weight2,	\
 			       pixman_fixed_frac (vx), unit_x, 0, FALSE);			\
 		dst += left_tz;									\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += left_tz;								\
 		vx += left_tz * unit_x;								\
 	    }											\
 	    if (width > 0)									\
 	    {											\
-		scanline_func (dst, mask,							\
-			       src1, src2, width, weight1, weight2, vx, unit_x, 0, FALSE);	\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, src1, src2, width, weight1, weight2,      \
+                               vx, unit_x, 0, FALSE);                                           \
 		dst += width;									\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += width;								\
 		vx += width * unit_x;								\
 	    }											\
 	    if (right_tz > 0)									\
 	    {											\
 		buf1[0] = src1[src_image->bits.width - 1];					\
 		buf1[1] = 0;									\
 		buf2[0] = src2[src_image->bits.width - 1];					\
 		buf2[1] = 0;									\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, right_tz, weight1, weight2,				\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, right_tz, weight1, weight2,   \
 			       pixman_fixed_frac (vx), unit_x, 0, FALSE);			\
 		dst += right_tz;								\
 		if (flags & FLAG_HAVE_NON_SOLID_MASK)						\
 		    mask += right_tz;								\
 	    }											\
 	    if (right_pad > 0)									\
 	    {											\
 		buf1[0] = buf1[1] = 0;								\
 		buf2[0] = buf2[1] = 0;								\
-		scanline_func (dst, mask,							\
-			       buf1, buf2, right_pad, weight1, weight2, 0, 0, 0, TRUE);		\
+		scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,	\
+			       scanline_buffer, mask, buf1, buf2, right_pad, weight1, weight2,  \
+                               0, 0, 0, TRUE);	                                                \
 	    }											\
 	}											\
 	else if (PIXMAN_REPEAT_ ## repeat_mode == PIXMAN_REPEAT_NORMAL)				\
 	{											\
 	    int32_t	    num_pixels;								\
 	    int32_t	    width_remain;							\
 	    src_type_t *    src_line_top;							\
 	    src_type_t *    src_line_bottom;							\
@@ -1120,17 +1167,18 @@ fast_composite_scaled_bilinear ## scale_
 		     * vx is in range [0, src_width_fixed - pixman_fixed_e]			\
 		     * So we are safe from overflow.						\
 		     */										\
 		    num_pixels = ((src_width_fixed - vx - pixman_fixed_e) / unit_x) + 1;	\
 												\
 		    if (num_pixels > width_remain)						\
 			num_pixels = width_remain;						\
 												\
-		    scanline_func (dst, mask, buf1, buf2, num_pixels,				\
+		    scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func,    \
+                                   dst, scanline_buffer, mask, buf1, buf2, num_pixels,          \
 				   weight1, weight2, pixman_fixed_frac(vx),			\
 				   unit_x, src_width_fixed, FALSE);				\
 												\
 		    width_remain -= num_pixels;							\
 		    vx += num_pixels * unit_x;							\
 		    dst += num_pixels;								\
 												\
 		    if (flags & FLAG_HAVE_NON_SOLID_MASK)					\
@@ -1149,41 +1197,47 @@ fast_composite_scaled_bilinear ## scale_
 		     * So we are safe from overflow here.					\
 		     */										\
 		    num_pixels = ((src_width_fixed - pixman_fixed_1 - vx - pixman_fixed_e)	\
 				  / unit_x) + 1;						\
 												\
 		    if (num_pixels > width_remain)						\
 			num_pixels = width_remain;						\
 												\
-		    scanline_func (dst, mask, src_line_top, src_line_bottom, num_pixels,	\
-				   weight1, weight2, vx, unit_x, src_width_fixed, FALSE);	\
+		    scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func,    \
+                                   dst, scanline_buffer, mask, src_line_top, src_line_bottom,   \
+                                   num_pixels, weight1, weight2, vx, unit_x, src_width_fixed,   \
+                                   FALSE);	                                                \
 												\
 		    width_remain -= num_pixels;							\
 		    vx += num_pixels * unit_x;							\
 		    dst += num_pixels;								\
 												\
 		    if (flags & FLAG_HAVE_NON_SOLID_MASK)					\
 		        mask += num_pixels;							\
 		}										\
 	    }											\
 	}											\
 	else											\
 	{											\
-	    scanline_func (dst, mask, src_first_line + src_stride * y1,				\
+	    scanline_func (dst_type_t, mask_type_t, src_type_t, fetch_func, op_func, dst,       \
+                           scanline_buffer, mask,                                               \
+                           src_first_line + src_stride * y1,					\
 			   src_first_line + src_stride * y2, width,				\
 			   weight1, weight2, vx, unit_x, max_vx, FALSE);			\
 	}											\
     }												\
+    if (scanline_buffer != (uint8_t *) stack_scanline_buffer)                                   \
+	free (scanline_buffer);                                                                 \
 }
 
 /* A workaround for old sun studio, see: https://bugs.freedesktop.org/show_bug.cgi?id=32764 */
-#define FAST_BILINEAR_MAINLOOP_COMMON(scale_func_name, scanline_func, src_type_t, mask_type_t,	\
+#define FAST_BILINEAR_MAINLOOP_COMMON(scale_func_name, fetch_func, op_func, src_type_t, mask_type_t,\
 				  dst_type_t, repeat_mode, flags)				\
-	FAST_BILINEAR_MAINLOOP_INT(_ ## scale_func_name, scanline_func, src_type_t, mask_type_t,\
+	FAST_BILINEAR_MAINLOOP_INT(_ ## scale_func_name, fetch_func, op_func, src_type_t, mask_type_t,\
 				  dst_type_t, repeat_mode, flags)
 
 #define SCALED_BILINEAR_FLAGS						\
     (FAST_PATH_SCALE_TRANSFORM	|					\
      FAST_PATH_NO_ALPHA_MAP	|					\
      FAST_PATH_BILINEAR_FILTER	|					\
      FAST_PATH_NO_ACCESSORS	|					\
      FAST_PATH_NARROW_FORMAT)
diff --git a/gfx/cairo/libpixman/src/pixman-sse2.c b/gfx/cairo/libpixman/src/pixman-sse2.c
--- a/gfx/cairo/libpixman/src/pixman-sse2.c
+++ b/gfx/cairo/libpixman/src/pixman-sse2.c
@@ -5404,30 +5404,33 @@ scaled_bilinear_scanline_sse2_8888_8888_
     if (w & 1)
     {
 	BILINEAR_INTERPOLATE_ONE_PIXEL (pix1);
 	*dst = pix1;
     }
 
 }
 
+/* Add extra NULL argument to the existing bilinear fast paths to indicate
+ * that we don't need two-pass processing */
+
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_cover_SRC,
-			       scaled_bilinear_scanline_sse2_8888_8888_SRC,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       COVER, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_pad_SRC,
-			       scaled_bilinear_scanline_sse2_8888_8888_SRC,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       PAD, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_none_SRC,
-			       scaled_bilinear_scanline_sse2_8888_8888_SRC,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NONE, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_normal_SRC,
-			       scaled_bilinear_scanline_sse2_8888_8888_SRC,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NORMAL, FLAG_NONE)
 
 static force_inline void
 scaled_bilinear_scanline_sse2_8888_8888_OVER (uint32_t *       dst,
 					      const uint32_t * mask,
 					      const uint32_t * src_top,
 					      const uint32_t * src_bottom,
@@ -5505,32 +5508,66 @@ scaled_bilinear_scanline_sse2_8888_8888_
 	}
 
 	w--;
 	dst++;
     }
 }
 
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_cover_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       COVER, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_pad_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       PAD, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_none_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NONE, FLAG_NONE)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8888_normal_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_OVER, NULL,
 			       uint32_t, uint32_t, uint32_t,
 			       NORMAL, FLAG_NONE)
 
+
+/* An example of SSE2 two-stage bilinear_over_8888_0565 fast path, which is implemented
+   as scaled_bilinear_scanline_sse2_8888_8888_SRC + op_bilinear_over_8888_0565 */
+
+void op_bilinear_over_8888_0565(uint16_t *dst, const uint32_t *mask, const uint32_t *src, int width)
+{
+    /* Note: this is not really fast and should be based on 8 pixel loop from sse2_composite_over_8888_0565 */
+    while (--width >= 0)
+    {
+	*dst = composite_over_8888_0565pixel (*src, *dst);
+	src++;
+	dst++;
+    }
+}
+
+FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_0565_cover_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       COVER, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_0565_pad_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       PAD, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_0565_none_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       NONE, FLAG_NONE)
+FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_0565_normal_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8888_SRC, op_bilinear_over_8888_0565,
+			       uint32_t, uint32_t, uint16_t,
+			       NORMAL, FLAG_NONE)
+
+/*****************************/
+
 static force_inline void
 scaled_bilinear_scanline_sse2_8888_8_8888_OVER (uint32_t *       dst,
 						const uint8_t  * mask,
 						const uint32_t * src_top,
 						const uint32_t * src_bottom,
 						int32_t          w,
 						int              wt,
 						int              wb,
@@ -5669,29 +5706,29 @@ scaled_bilinear_scanline_sse2_8888_8_888
 	}
 
 	w--;
 	dst++;
     }
 }
 
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8_8888_cover_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER, NULL,
 			       uint32_t, uint8_t, uint32_t,
 			       COVER, FLAG_HAVE_NON_SOLID_MASK)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8_8888_pad_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER, NULL,
 			       uint32_t, uint8_t, uint32_t,
 			       PAD, FLAG_HAVE_NON_SOLID_MASK)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8_8888_none_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER, NULL,
 			       uint32_t, uint8_t, uint32_t,
 			       NONE, FLAG_HAVE_NON_SOLID_MASK)
 FAST_BILINEAR_MAINLOOP_COMMON (sse2_8888_8_8888_normal_OVER,
-			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER,
+			       scaled_bilinear_scanline_sse2_8888_8_8888_OVER, NULL,
 			       uint32_t, uint8_t, uint32_t,
 			       NORMAL, FLAG_HAVE_NON_SOLID_MASK)
 
 static const pixman_fast_path_t sse2_fast_paths[] =
 {
     /* PIXMAN_OP_OVER */
     PIXMAN_STD_FAST_PATH (OVER, solid, a8, r5g6b5, sse2_composite_over_n_8_0565),
     PIXMAN_STD_FAST_PATH (OVER, solid, a8, b5g6r5, sse2_composite_over_n_8_0565),
@@ -5808,16 +5845,21 @@ static const pixman_fast_path_t sse2_fas
     SIMPLE_BILINEAR_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, sse2_8888_8888),
     SIMPLE_BILINEAR_FAST_PATH (OVER, a8b8g8r8, a8b8g8r8, sse2_8888_8888),
 
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8r8g8b8, x8r8g8b8, sse2_8888_8_8888),
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8b8g8r8, x8b8g8r8, sse2_8888_8_8888),
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8r8g8b8, a8r8g8b8, sse2_8888_8_8888),
     SIMPLE_BILINEAR_A8_MASK_FAST_PATH (OVER, a8b8g8r8, a8b8g8r8, sse2_8888_8_8888),
 
+    /* and here the needed entries are added to the fast path table */
+
+    SIMPLE_BILINEAR_FAST_PATH (OVER, a8r8g8b8, r5g6b5, sse2_8888_0565),
+    SIMPLE_BILINEAR_FAST_PATH (OVER, a8b8g8r8, b5g6r5, sse2_8888_0565),
+
     { PIXMAN_OP_NONE },
 };
 
 static pixman_bool_t
 sse2_blt (pixman_implementation_t *imp,
           uint32_t *               src_bits,
           uint32_t *               dst_bits,
           int                      src_stride,