path: root/recipes/xorg-lib/pixman/remove-broken.patch

From: Siarhei Siamashka <siarhei.siamashka@nokia.com>
Date: Sun, 26 Jul 2009 22:21:26 +0000 (+0300)
Subject: ARM: Removal of unused/broken NEON code
X-Git-Url: http://siarhei.siamashka.name/gitweb/?p=pixman.git;a=commitdiff_plain;h=7ef2322eefcccc28a2d45c0da22c0fee88b8f464

ARM: Removal of unused/broken NEON code
---

diff --git a/pixman/pixman-arm-neon.c b/pixman/pixman-arm-neon.c
index 4125d1b..9404c70 100644
--- a/pixman/pixman-arm-neon.c
+++ b/pixman/pixman-arm-neon.c
@@ -1895,710 +1895,6 @@ pixman_fill_neon (uint32_t *bits,
 #endif
 }
 
-/* TODO: is there a more generic way of doing this being introduced? */
-#define NEON_SCANLINE_BUFFER_PIXELS (1024)
-
-static inline void
-neon_quadword_copy (void *   dst,
-		    void *   src,
-		    uint32_t count,         /* of quadwords */
-		    uint32_t trailer_count  /* of bytes */)
-{
-    uint8_t *t_dst = dst, *t_src = src;
-
-    /* Uses aligned multi-register loads to maximise read bandwidth
-     * on uncached memory such as framebuffers
-     * The accesses do not have the aligned qualifiers, so that the copy
-     * may convert between aligned-uncached and unaligned-cached memory.
-     * It is assumed that the CPU can infer alignedness from the address.
-     */
-
-#ifdef USE_GCC_INLINE_ASM
-
-    asm volatile (
-        "	cmp       %[count], #8				\n"
-        "	blt 1f    @ skip oversized fragments		\n"
-        "0: @ start with eight quadwords at a time		\n"
-        "	sub       %[count], %[count], #8		\n"
-        "	vld1.8    {d16, d17, d18, d19}, [%[src]]!		\n"
-        "	vld1.8    {d20, d21, d22, d23}, [%[src]]!		\n"
-        "	vld1.8    {d24, d25, d26, d27}, [%[src]]!		\n"
-        "	vld1.8    {d28, d29, d30, d31}, [%[src]]!		\n"
-        "	cmp       %[count], #8				\n"
-        "	vst1.8    {d16, d17, d18, d19}, [%[dst]]!		\n"
-        "	vst1.8    {d20, d21, d22, d23}, [%[dst]]!		\n"
-        "	vst1.8    {d24, d25, d26, d27}, [%[dst]]!		\n"
-        "	vst1.8    {d28, d29, d30, d31}, [%[dst]]!		\n"
-        "	bge 0b						\n"
-        "1: @ four quadwords					\n"
-        "	tst       %[count], #4				\n"
-        "	beq 2f    @ skip oversized fragment		\n"
-        "	vld1.8    {d16, d17, d18, d19}, [%[src]]!		\n"
-        "	vld1.8    {d20, d21, d22, d23}, [%[src]]!		\n"
-        "	vst1.8    {d16, d17, d18, d19}, [%[dst]]!		\n"
-        "	vst1.8    {d20, d21, d22, d23}, [%[dst]]!		\n"
-        "2: @ two quadwords					\n"
-        "	tst       %[count], #2				\n"
-        "	beq 3f    @ skip oversized fragment		\n"
-        "	vld1.8    {d16, d17, d18, d19}, [%[src]]!		\n"
-        "	vst1.8    {d16, d17, d18, d19}, [%[dst]]!		\n"
-        "3: @ one quadword					\n"
-        "	tst       %[count], #1				\n"
-        "	beq 4f    @ skip oversized fragment		\n"
-        "	vld1.8    {d16, d17}, [%[src]]!			\n"
-        "	vst1.8    {d16, d17}, [%[dst]]!			\n"
-        "4: @ end						\n"
-
-        /* Clobbered input registers marked as input/outputs */
-	: [dst] "+r" (t_dst), [src] "+r" (t_src), [count] "+r" (count)
-
-	  /* No unclobbered inputs */
-	:
-
-        /* Clobbered vector registers */
-	: "d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23", "d24", "d25",
-	  "d26", "d27", "d28", "d29", "d30", "d31", "cc", "memory");
-
-#else
-
-    while (count >= 8)
-    {
-	uint8x16x4_t t1 = vld4q_u8 (t_src);
-	uint8x16x4_t t2 = vld4q_u8 (t_src + sizeof(uint8x16x4_t));
-	
-	t_src += sizeof(uint8x16x4_t) * 2;
-	vst4q_u8 (t_dst, t1);
-	vst4q_u8 (t_dst + sizeof(uint8x16x4_t), t2);
-	t_dst += sizeof(uint8x16x4_t) * 2;
-	count -= 8;
-    }
-
-    if (count & 4)
-    {
-	uint8x16x4_t t1 = vld4q_u8 (t_src);
-	
-	t_src += sizeof(uint8x16x4_t);
-	vst4q_u8 (t_dst, t1);
-	t_dst += sizeof(uint8x16x4_t);
-    }
-
-    if (count & 2)
-    {
-	uint8x8x4_t t1 = vld4_u8 (t_src);
-	
-	t_src += sizeof(uint8x8x4_t);
-	vst4_u8 (t_dst, t1);
-	t_dst += sizeof(uint8x8x4_t);
-    }
-
-    if (count & 1)
-    {
-	uint8x16_t t1 = vld1q_u8 (t_src);
-	
-	t_src += sizeof(uint8x16_t);
-	vst1q_u8 (t_dst, t1);
-	t_dst += sizeof(uint8x16_t);
-    }
-
-#endif  /* !USE_GCC_INLINE_ASM */
-
-    if (trailer_count)
-    {
-	if (trailer_count & 8)
-	{
-	    uint8x8_t t1 = vld1_u8 (t_src);
-	    
-	    t_src += sizeof(uint8x8_t);
-	    vst1_u8 (t_dst, t1);
-	    t_dst += sizeof(uint8x8_t);
-	}
-
-	if (trailer_count & 4)
-	{
-	    *((uint32_t*) t_dst) = *((uint32_t*) t_src);
-	    
-	    t_dst += 4;
-	    t_src += 4;
-	}
-
-	if (trailer_count & 2)
-	{
-	    *((uint16_t*) t_dst) = *((uint16_t*) t_src);
-	    
-	    t_dst += 2;
-	    t_src += 2;
-	}
-
-	if (trailer_count & 1)
-	{
-	    *t_dst++ = *t_src++;
-	}
-    }
-}
-
-static inline void
-solid_over_565_8_pix_neon (uint32_t  glyph_colour,
-                           uint16_t *dest,
-                           uint8_t * in_mask,
-                           uint32_t  dest_stride,    /* bytes, not elements */
-                           uint32_t  mask_stride,
-                           uint32_t  count           /* 8-pixel groups */)
-{
-    /* Inner loop of glyph blitter (solid colour, alpha mask) */
-
-#ifdef USE_GCC_INLINE_ASM
-
-    asm volatile (
-        "	vld4.8 {d20[], d21[], d22[], d23[]}, [%[glyph_colour]]  @ splat solid colour components	\n"
-        "0:	@ loop																				\n"
-        "	vld1.16   {d0, d1}, [%[dest]]         @ load first pixels from framebuffer			\n"
-        "	vld1.8    {d17}, [%[in_mask]]         @ load alpha mask of glyph						\n"
-        "	vmull.u8  q9, d17, d23               @ apply glyph colour alpha to mask				\n"
-        "	vshrn.u16 d17, q9, #8                @ reformat it to match original mask			\n"
-        "	vmvn      d18, d17                   @ we need the inverse mask for the background	\n"
-        "	vsli.u16  q3, q0, #5                 @ duplicate framebuffer blue bits				\n"
-        "	vshrn.u16 d2, q0, #8                 @ unpack red from framebuffer pixels			\n"
-        "	vshrn.u16 d4, q0, #3                 @ unpack green									\n"
-        "	vsri.u8   d2, d2, #5                 @ duplicate red bits (extend 5 to 8)			\n"
-        "	vshrn.u16 d6, q3, #2                 @ unpack extended blue (truncate 10 to 8)		\n"
-        "	vsri.u8   d4, d4, #6                 @ duplicate green bits (extend 6 to 8)			\n"
-        "	vmull.u8  q1, d2, d18                @ apply inverse mask to background red...		\n"
-        "	vmull.u8  q2, d4, d18                @ ...green...									\n"
-        "	vmull.u8  q3, d6, d18                @ ...blue										\n"
-        "	subs      %[count], %[count], #1     @ decrement/test loop counter					\n"
-        "	vmlal.u8  q1, d17, d22               @ add masked foreground red...					\n"
-        "	vmlal.u8  q2, d17, d21               @ ...green...									\n"
-        "	vmlal.u8  q3, d17, d20               @ ...blue										\n"
-        "	add %[in_mask], %[in_mask], %[mask_stride] @ advance mask pointer, while we wait		\n"
-        "	vsri.16   q1, q2, #5                 @ pack green behind red						\n"
-        "	vsri.16   q1, q3, #11                @ pack blue into pixels						\n"
-        "	vst1.16   {d2, d3}, [%[dest]]         @ store composited pixels						\n"
-        "	add %[dest], %[dest], %[dest_stride]  @ advance framebuffer pointer					\n"
-        "	bne 0b                               @ next please									\n"
-
-	/* Clobbered registers marked as input/outputs */
-	: [dest] "+r" (dest), [in_mask] "+r" (in_mask), [count] "+r" (count)
-	  
-	  /* Inputs */
-	: [dest_stride] "r" (dest_stride), [mask_stride] "r" (mask_stride), [glyph_colour] "r" (&glyph_colour)
-
-	  /* Clobbers, including the inputs we modify, and potentially lots of memory */
-	: "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d17", "d18", "d19",
-	  "d20", "d21", "d22", "d23", "d24", "d25", "cc", "memory"
-        );
-
-#else
-
-    uint8x8x4_t solid_colour = vld4_dup_u8 ((uint8_t*) &glyph_colour);
-
-    while (count--)
-    {
-	uint16x8_t pixels = vld1q_u16 (dest);
-	uint8x8_t mask = vshrn_n_u16 (vmull_u8 (solid_colour.val[3], vld1_u8 (in_mask)), 8);
-	uint8x8_t mask_image = vmvn_u8 (mask);
-
-	uint8x8_t t_red   = vshrn_n_u16 (pixels, 8);
-	uint8x8_t t_green = vshrn_n_u16 (pixels, 3);
-	uint8x8_t t_blue  = vshrn_n_u16 (vsli_n_u8 (pixels, pixels, 5), 2);
-
-	uint16x8_t s_red   = vmull_u8 (vsri_n_u8 (t_red, t_red, 5), mask_image);
-	uint16x8_t s_green = vmull_u8 (vsri_n_u8 (t_green, t_green, 6), mask_image);
-	uint16x8_t s_blue  = vmull_u8 (t_blue, mask_image);
-
-	s_red   = vmlal (s_red, mask, solid_colour.val[2]);
-	s_green = vmlal (s_green, mask, solid_colour.val[1]);
-	s_blue  = vmlal (s_blue, mask, solid_colour.val[0]);
-
-	pixels = vsri_n_u16 (s_red, s_green, 5);
-	pixels = vsri_n_u16 (pixels, s_blue, 11);
-	vst1q_u16 (dest, pixels);
-
-	dest += dest_stride;
-	mask += mask_stride;
-    }
-
-#endif
-}
-
-#if 0 /* this is broken currently */
-static void
-neon_composite_over_n_8_0565 (pixman_implementation_t * impl,
-                              pixman_op_t               op,
-                              pixman_image_t *          src_image,
-                              pixman_image_t *          mask_image,
-                              pixman_image_t *          dst_image,
-                              int32_t                   src_x,
-                              int32_t                   src_y,
-                              int32_t                   mask_x,
-                              int32_t                   mask_y,
-                              int32_t                   dest_x,
-                              int32_t                   dest_y,
-                              int32_t                   width,
-                              int32_t                   height)
-{
-    uint32_t  src, srca;
-    uint16_t *dst_line, *aligned_line;
-    uint8_t  *mask_line;
-    uint32_t  dst_stride, mask_stride;
-    uint32_t  kernel_count, copy_count, copy_tail;
-    uint8_t   kernel_offset, copy_offset;
-
-    src = _pixman_image_get_solid (src_image, dst_image->bits.format);
-
-    /* bail out if fully transparent or degenerate */
-    srca = src >> 24;
-    if (src == 0)
-	return;
-
-    if (width == 0 || height == 0)
-	return;
-
-    if (width > NEON_SCANLINE_BUFFER_PIXELS)
-    {
-	/* split the blit, so we can use a fixed-size scanline buffer
-	 * TODO: there must be a more elegant way of doing this.
-	 */
-	int x;
-	for (x = 0; x < width; x += NEON_SCANLINE_BUFFER_PIXELS)
-	{
-	    neon_composite_over_n_8_0565 (
-		impl, op,
-		src_image, mask_image, dst_image,
-		src_x + x, src_y, mask_x + x, mask_y, dest_x + x, dest_y,
-		(x + NEON_SCANLINE_BUFFER_PIXELS > width) ? width - x : NEON_SCANLINE_BUFFER_PIXELS, height);
-	}
-
-	return;
-    }
-    
-    PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, uint16_t, dst_stride, dst_line, 1);
-    PIXMAN_IMAGE_GET_LINE (mask_image, mask_x, mask_y, uint8_t, mask_stride, mask_line, 1);
-
-    /* keep within minimum number of aligned quadwords on width
-     * while also keeping the minimum number of columns to process
-     */
-    {
-	unsigned long aligned_left = (unsigned long)(dst_line) & ~0xF;
-	unsigned long aligned_right = (((unsigned long)(dst_line + width)) + 0xF) & ~0xF;
-	unsigned long ceiling_length = (((unsigned long) width) * sizeof(*dst_line) + 0xF) & ~0xF;
-
-	/* the fast copy should be quadword aligned */
-	copy_offset = dst_line - ((uint16_t*) aligned_left);
-	aligned_line = dst_line - copy_offset;
-	copy_count = (uint32_t) ((aligned_right - aligned_left) >> 4);
-	copy_tail = 0;
-
-	if (aligned_right - aligned_left > ceiling_length)
-	{
-	    /* unaligned routine is tightest */
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    kernel_offset = copy_offset;
-	}
-	else
-	{
-	    /* aligned routine is equally tight, so it is safer to align */
-	    kernel_count = copy_count;
-	    kernel_offset = 0;
-	}
-
-	/* We should avoid reading beyond scanline ends for safety */
-	if (aligned_line < (dst_line - dest_x) ||
-	    (aligned_line + (copy_count * 16 / sizeof(*dst_line))) > ((dst_line - dest_x) + dst_image->bits.width))
-	{
-	    /* switch to precise read */
-	    copy_offset = kernel_offset = 0;
-	    aligned_line = dst_line;
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    copy_count = (width * sizeof(*dst_line)) >> 4;
-	    copy_tail = (width * sizeof(*dst_line)) & 0xF;
-	}
-    }
-
-    {
-	uint16_t scan_line[NEON_SCANLINE_BUFFER_PIXELS + 8];         /* deliberately not initialised */
-	uint8_t glyph_line[NEON_SCANLINE_BUFFER_PIXELS + 8];
-	int y = height;
-
-	/* row-major order */
-	/* left edge, middle block, right edge */
-	for ( ; y--; mask_line += mask_stride, aligned_line += dst_stride, dst_line += dst_stride)
-	{
-	    /* We don't want to overrun the edges of the glyph,
-	     * so realign the edge data into known buffers
-	     */
-	    neon_quadword_copy (glyph_line + copy_offset, mask_line, width >> 4, width & 0xF);
-
-	    /* Uncached framebuffer access is really, really slow
-	     * if we do it piecemeal. It should be much faster if we
-	     * grab it all at once. One scanline should easily fit in
-	     * L1 cache, so this should not waste RAM bandwidth.
-	     */
-	    neon_quadword_copy (scan_line, aligned_line, copy_count, copy_tail);
-
-	    /* Apply the actual filter */
-	    solid_over_565_8_pix_neon (
-		src, scan_line + kernel_offset,
-		glyph_line + kernel_offset, 8 * sizeof(*dst_line),
-		8, kernel_count);
-
-	    /* Copy the modified scanline back */
-	    neon_quadword_copy (dst_line, scan_line + copy_offset,
-				width >> 3, (width & 7) * 2);
-	}
-    }
-}
-#endif
-
-#ifdef USE_GCC_INLINE_ASM
-
-static inline void
-plain_over_565_8_pix_neon (uint32_t  colour,
-			   uint16_t *dest,
-			   uint32_t  dest_stride,     /* bytes, not elements */
-			   uint32_t  count            /* 8-pixel groups */)
-{
-    /* Inner loop for plain translucent rects
-     * (solid colour without alpha mask)
-     */
-    asm volatile (
-        "	vld4.8   {d20[], d21[], d22[], d23[]}, [%[colour]]  @ solid colour load/splat \n"
-        "	vmull.u8  q12, d23, d22              @ premultiply alpha red   \n"
-        "	vmull.u8  q13, d23, d21              @ premultiply alpha green \n"
-        "	vmull.u8  q14, d23, d20              @ premultiply alpha blue  \n"
-        "	vmvn      d18, d23                   @ inverse alpha for background \n"
-        "0:	@ loop\n"
-        "	vld1.16   {d0, d1}, [%[dest]]         @ load first pixels from framebuffer	\n"
-        "	vshrn.u16 d2, q0, #8                 @ unpack red from framebuffer pixels	\n"
-        "	vshrn.u16 d4, q0, #3                 @ unpack green				\n"
-        "	vsli.u16  q3, q0, #5                 @ duplicate framebuffer blue bits		\n"
-        "	vsri.u8   d2, d2, #5                 @ duplicate red bits (extend 5 to 8)	\n"
-        "	vsri.u8   d4, d4, #6                 @ duplicate green bits (extend 6 to 8)	\n"
-        "	vshrn.u16 d6, q3, #2                 @ unpack extended blue (truncate 10 to 8)	\n"
-        "	vmov      q0, q12                    @ retrieve foreground red   \n"
-        "	vmlal.u8  q0, d2, d18                @ blend red - my kingdom for a four-operand MLA \n"
-        "	vmov      q1, q13                    @ retrieve foreground green \n"
-        "	vmlal.u8  q1, d4, d18                @ blend green               \n"
-        "	vmov      q2, q14                    @ retrieve foreground blue  \n"
-        "	vmlal.u8  q2, d6, d18                @ blend blue                \n"
-        "	subs      %[count], %[count], #1     @ decrement/test loop counter		\n"
-        "	vsri.16   q0, q1, #5                 @ pack green behind red			\n"
-        "	vsri.16   q0, q2, #11                @ pack blue into pixels			\n"
-        "	vst1.16   {d0, d1}, [%[dest]]         @ store composited pixels			\n"
-        "	add %[dest], %[dest], %[dest_stride]  @ advance framebuffer pointer		\n"
-        "	bne 0b                               @ next please				\n"
-
-        /* Clobbered registers marked as input/outputs */
-	: [dest] "+r" (dest), [count] "+r" (count)
-
-	  /* Inputs */
-	: [dest_stride] "r" (dest_stride), [colour] "r" (&colour)
-
-	  /* Clobbers, including the inputs we modify, and
-	   * potentially lots of memory
-	   */
-	: "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d18", "d19",
-	  "d20", "d21", "d22", "d23", "d24", "d25", "d26", "d27", "d28", "d29",
-	  "cc", "memory"
-        );
-}
-
-static void
-neon_composite_over_n_0565 (pixman_implementation_t * impl,
-                            pixman_op_t               op,
-                            pixman_image_t *          src_image,
-                            pixman_image_t *          mask_image,
-                            pixman_image_t *          dst_image,
-                            int32_t                   src_x,
-                            int32_t                   src_y,
-                            int32_t                   mask_x,
-                            int32_t                   mask_y,
-                            int32_t                   dest_x,
-                            int32_t                   dest_y,
-                            int32_t                   width,
-                            int32_t                   height)
-{
-    uint32_t src, srca;
-    uint16_t    *dst_line, *aligned_line;
-    uint32_t dst_stride;
-    uint32_t kernel_count, copy_count, copy_tail;
-    uint8_t kernel_offset, copy_offset;
-
-    src = _pixman_image_get_solid (src_image, dst_image->bits.format);
-
-    /* bail out if fully transparent */
-    srca = src >> 24;
-    if (src == 0)
-	return;
-    
-    if (width == 0 || height == 0)
-	return;
-
-    if (width > NEON_SCANLINE_BUFFER_PIXELS)
-    {
-	/* split the blit, so we can use a fixed-size scanline buffer *
-	 * TODO: there must be a more elegant way of doing this.
-	 */
-	int x;
-	
-	for (x = 0; x < width; x += NEON_SCANLINE_BUFFER_PIXELS)
-	{
-	    neon_composite_over_n_0565 (
-		impl, op,
-		src_image, mask_image, dst_image,
-		src_x + x, src_y, mask_x + x, mask_y, dest_x + x, dest_y,
-		(x + NEON_SCANLINE_BUFFER_PIXELS > width) ? width - x : NEON_SCANLINE_BUFFER_PIXELS, height);
-	}
-	return;
-    }
-
-    PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, uint16_t, dst_stride, dst_line, 1);
-
-    /* keep within minimum number of aligned quadwords on width
-     * while also keeping the minimum number of columns to process
-     */
-    {
-	unsigned long aligned_left = (unsigned long)(dst_line) & ~0xF;
-	unsigned long aligned_right = (((unsigned long)(dst_line + width)) + 0xF) & ~0xF;
-	unsigned long ceiling_length = (((unsigned long) width) * sizeof(*dst_line) + 0xF) & ~0xF;
-
-	/* the fast copy should be quadword aligned */
-	copy_offset = dst_line - ((uint16_t*) aligned_left);
-	aligned_line = dst_line - copy_offset;
-	copy_count = (uint32_t) ((aligned_right - aligned_left) >> 4);
-	copy_tail = 0;
-
-	if (aligned_right - aligned_left > ceiling_length)
-	{
-	    /* unaligned routine is tightest */
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    kernel_offset = copy_offset;
-	}
-	else
-	{
-	    /* aligned routine is equally tight, so it is safer to align */
-	    kernel_count = copy_count;
-	    kernel_offset = 0;
-	}
-
-	/* We should avoid reading beyond scanline ends for safety */
-	if (aligned_line < (dst_line - dest_x) ||
-	    (aligned_line + (copy_count * 16 / sizeof(*dst_line))) > ((dst_line - dest_x) + dst_image->bits.width))
-	{
-	    /* switch to precise read */
-	    copy_offset = kernel_offset = 0;
-	    aligned_line = dst_line;
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    copy_count = (width * sizeof(*dst_line)) >> 4;
-	    copy_tail = (width * sizeof(*dst_line)) & 0xF;
-	}
-    }
-
-    {
-	uint16_t scan_line[NEON_SCANLINE_BUFFER_PIXELS + 8];  /* deliberately not initialised */
-
-	/* row-major order */
-	/* left edge, middle block, right edge */
-	for ( ; height--; aligned_line += dst_stride, dst_line += dst_stride)
-	{
-	    /* Uncached framebuffer access is really, really slow if we do it piecemeal.
-	     * It should be much faster if we grab it all at once.
-	     * One scanline should easily fit in L1 cache, so this should
-	     * not waste RAM bandwidth.
-	     */
-	    neon_quadword_copy (scan_line, aligned_line, copy_count, copy_tail);
-
-	    /* Apply the actual filter */
-	    plain_over_565_8_pix_neon (
-		src, scan_line + kernel_offset, 8 * sizeof(*dst_line), kernel_count);
-
-	    /* Copy the modified scanline back */
-	    neon_quadword_copy (
-		dst_line, scan_line + copy_offset, width >> 3, (width & 7) * 2);
-	}
-    }
-}
-
-static inline void
-ARGB8_over_565_8_pix_neon (uint32_t *src,
-                           uint16_t *dest,
-                           uint32_t  src_stride,     /* bytes, not elements */
-                           uint32_t  count           /* 8-pixel groups */)
-{
-    asm volatile (
-        "0:	@ loop\n"
-        "	pld   [%[src], %[src_stride]]         @ preload from next scanline	\n"
-        "	vld1.16   {d0, d1}, [%[dest]]         @ load pixels from framebuffer	\n"
-        "	vld4.8   {d20, d21, d22, d23},[%[src]]! @ load source image pixels		\n"
-        "	vsli.u16  q3, q0, #5                 @ duplicate framebuffer blue bits		\n"
-        "	vshrn.u16 d2, q0, #8                 @ unpack red from framebuffer pixels	\n"
-        "	vshrn.u16 d4, q0, #3                 @ unpack green				\n"
-        "	vmvn      d18, d23                   @ we need the inverse alpha for the background	\n"
-        "	vsri.u8   d2, d2, #5                 @ duplicate red bits (extend 5 to 8)	\n"
-        "	vshrn.u16 d6, q3, #2                 @ unpack extended blue (truncate 10 to 8)	\n"
-        "	vsri.u8   d4, d4, #6                 @ duplicate green bits (extend 6 to 8)	\n"
-        "	vmull.u8  q1, d2, d18                @ apply inverse alpha to background red...	\n"
-        "	vmull.u8  q2, d4, d18                @ ...green...				\n"
-        "	vmull.u8  q3, d6, d18                @ ...blue					\n"
-        "	subs      %[count], %[count], #1     @ decrement/test loop counter		\n"
-        "	vmlal.u8  q1, d23, d22               @ add blended foreground red...		\n"
-        "	vmlal.u8  q2, d23, d21               @ ...green...				\n"
-        "	vmlal.u8  q3, d23, d20               @ ...blue					\n"
-        "	vsri.16   q1, q2, #5                 @ pack green behind red			\n"
-        "	vsri.16   q1, q3, #11                @ pack blue into pixels			\n"
-        "	vst1.16   {d2, d3}, [%[dest]]!        @ store composited pixels			\n"
-        "	bne 0b                               @ next please				\n"
-
-        /* Clobbered registers marked as input/outputs */
-	: [dest] "+r" (dest), [src] "+r" (src), [count] "+r" (count)
-
-	  /* Inputs */
-	: [src_stride] "r" (src_stride)
-
-	  /* Clobbers, including the inputs we modify, and potentially lots of memory */
-	: "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d17", "d18", "d20",
-	  "d21", "d22", "d23", "cc", "memory"
-        );
-}
-
-static void
-neon_composite_over_8888_0565 (pixman_implementation_t * impl,
-                               pixman_op_t               op,
-                               pixman_image_t *          src_image,
-                               pixman_image_t *          mask_image,
-                               pixman_image_t *          dst_image,
-                               int32_t                   src_x,
-                               int32_t                   src_y,
-                               int32_t                   mask_x,
-                               int32_t                   mask_y,
-                               int32_t                   dest_x,
-                               int32_t                   dest_y,
-                               int32_t                   width,
-                               int32_t                   height)
-{
-    uint32_t    *src_line;
-    uint16_t    *dst_line, *aligned_line;
-    uint32_t dst_stride, src_stride;
-    uint32_t kernel_count, copy_count, copy_tail;
-    uint8_t kernel_offset, copy_offset;
-
-    /* we assume mask is opaque 
-     * so the only alpha to deal with is embedded in src
-     */
-    if (width > NEON_SCANLINE_BUFFER_PIXELS)
-    {
-	/* split the blit, so we can use a fixed-size scanline buffer */
-	int x;
-	for (x = 0; x < width; x += NEON_SCANLINE_BUFFER_PIXELS)
-	{
-	    neon_composite_over_8888_0565 (
-		impl, op,
-		src_image, mask_image, dst_image,
-		src_x + x, src_y, mask_x + x, mask_y, dest_x + x, dest_y,
-		(x + NEON_SCANLINE_BUFFER_PIXELS > width) ? width - x : NEON_SCANLINE_BUFFER_PIXELS, height);
-	}
-	return;
-    }
-
-    PIXMAN_IMAGE_GET_LINE (dst_image, dest_x, dest_y, uint16_t, dst_stride, dst_line, 1);
-    PIXMAN_IMAGE_GET_LINE (src_image, src_x, src_y, uint32_t, src_stride, src_line, 1);
-
-    /* keep within minimum number of aligned quadwords on width
-     * while also keeping the minimum number of columns to process
-     */
-    {
-	unsigned long aligned_left = (unsigned long)(dst_line) & ~0xF;
-	unsigned long aligned_right = (((unsigned long)(dst_line + width)) + 0xF) & ~0xF;
-	unsigned long ceiling_length = (((unsigned long) width) * sizeof(*dst_line) + 0xF) & ~0xF;
-
-	/* the fast copy should be quadword aligned */
-	copy_offset = dst_line - ((uint16_t*) aligned_left);
-	aligned_line = dst_line - copy_offset;
-	copy_count = (uint32_t) ((aligned_right - aligned_left) >> 4);
-	copy_tail = 0;
-
-	if (aligned_right - aligned_left > ceiling_length)
-	{
-	    /* unaligned routine is tightest */
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    kernel_offset = copy_offset;
-	}
-	else
-	{
-	    /* aligned routine is equally tight, so it is safer to align */
-	    kernel_count = copy_count;
-	    kernel_offset = 0;
-	}
-
-	/* We should avoid reading beyond scanline ends for safety */
-	if (aligned_line < (dst_line - dest_x) ||
-	    (aligned_line + (copy_count * 16 / sizeof(*dst_line))) > ((dst_line - dest_x) + dst_image->bits.width))
-	{
-	    /* switch to precise read */
-	    copy_offset = kernel_offset = 0;
-	    aligned_line = dst_line;
-	    kernel_count = (uint32_t) (ceiling_length >> 4);
-	    copy_count = (width * sizeof(*dst_line)) >> 4;
-	    copy_tail = (width * sizeof(*dst_line)) & 0xF;
-	}
-    }
-
-    /* Preload the first input scanline */
-    {
-	uint8_t *src_ptr = (uint8_t*) src_line;
-	uint32_t count = (width + 15) / 16;
-
-#ifdef USE_GCC_INLINE_ASM
-	asm volatile (
-	    "0: @ loop						\n"
-	    "	subs    %[count], %[count], #1			\n"
-	    "	pld     [%[src]]				\n"
-	    "	add     %[src], %[src], #64			\n"
-	    "	bgt 0b						\n"
-
-	    /* Clobbered input registers marked as input/outputs */
-	    : [src] "+r" (src_ptr), [count] "+r" (count)
-	    :     /* no unclobbered inputs */
-	    : "cc"
-	    );
-#else
-	do
-	{
-	    __pld (src_ptr);
-	    src_ptr += 64;
-	}
-	while (--count);
-#endif
-    }
-
-    {
-	uint16_t scan_line[NEON_SCANLINE_BUFFER_PIXELS + 8]; /* deliberately not initialised */
-
-	/* row-major order */
-	/* left edge, middle block, right edge */
-	for ( ; height--; src_line += src_stride, aligned_line += dst_stride)
-	{
-	    /* Uncached framebuffer access is really, really slow if we do
-	     * it piecemeal. It should be much faster if we grab it all at
-	     * once. One scanline should easily fit in L1 cache, so this
-	     * should not waste RAM bandwidth.
-	     */
-	    neon_quadword_copy (scan_line, aligned_line, copy_count, copy_tail);
-
-	    /* Apply the actual filter */
-	    ARGB8_over_565_8_pix_neon (
-		src_line, scan_line + kernel_offset,
-		src_stride * sizeof(*src_line), kernel_count);
-
-	    /* Copy the modified scanline back */
-	    neon_quadword_copy (dst_line,
-				scan_line + copy_offset,
-				width >> 3, (width & 7) * 2);
-	}
-    }
-}
-
-#endif  /* USE_GCC_INLINE_ASM */
-
 static const pixman_fast_path_t arm_neon_fast_path_array[] =
 {
     { PIXMAN_OP_ADD,  PIXMAN_solid,    PIXMAN_a8,       PIXMAN_a8,       neon_composite_add_8888_8_8,     0 },
@@ -2612,12 +1908,6 @@ static const pixman_fast_path_t arm_neon_fast_path_array[] =
 #ifdef USE_GCC_INLINE_ASM
     { PIXMAN_OP_SRC,  PIXMAN_r5g6b5,   PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_src_16_16,        0 },
     { PIXMAN_OP_SRC,  PIXMAN_b5g6r5,   PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_src_16_16,        0 },
-#if 0 /* this code has some bugs */
-    { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_over_n_0565,      0 },
-    { PIXMAN_OP_OVER, PIXMAN_solid,    PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_over_n_0565,      0 },
-    { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_r5g6b5,   neon_composite_over_8888_0565,   0 },
-    { PIXMAN_OP_OVER, PIXMAN_a8b8g8r8, PIXMAN_null,     PIXMAN_b5g6r5,   neon_composite_over_8888_0565,   0 },
-#endif
 #endif
     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_a8r8g8b8, neon_composite_over_8888_8888,   0 },
     { PIXMAN_OP_OVER, PIXMAN_a8r8g8b8, PIXMAN_null,     PIXMAN_x8r8g8b8, neon_composite_over_8888_8888,   0 },
@@ -2668,79 +1958,6 @@ arm_neon_composite (pixman_implementation_t *imp,
 }
 
 static pixman_bool_t
-pixman_blt_neon (void *src_bits,
-                 void *dst_bits,
-                 int   src_stride,
-                 int   dst_stride,
-                 int   src_bpp,
-                 int   dst_bpp,
-                 int   src_x,
-                 int   src_y,
-                 int   dst_x,
-                 int   dst_y,
-                 int   width,
-                 int   height)
-{
-    if (!width || !height)
-	return TRUE;
-
-    /* accelerate only straight copies involving complete bytes */
-    if (src_bpp != dst_bpp || (src_bpp & 7))
-	return FALSE;
-
-    {
-	uint32_t bytes_per_pixel = src_bpp >> 3;
-	uint32_t byte_width = width * bytes_per_pixel;
-	/* parameter is in words for some reason */
-	int32_t src_stride_bytes = src_stride * 4;
-	int32_t dst_stride_bytes = dst_stride * 4;
-	uint8_t *src_bytes = ((uint8_t*) src_bits) +
-	    src_y * src_stride_bytes + src_x * bytes_per_pixel;
-	uint8_t *dst_bytes = ((uint8_t*) dst_bits) +
-	    dst_y * dst_stride_bytes + dst_x * bytes_per_pixel;
-	uint32_t quadword_count = byte_width / 16;
-	uint32_t offset         = byte_width % 16;
-
-	while (height--)
-	{
-	    neon_quadword_copy (dst_bytes, src_bytes, quadword_count, offset);
-	    src_bytes += src_stride_bytes;
-	    dst_bytes += dst_stride_bytes;
-	}
-    }
-
-    return TRUE;
-}
-
-static pixman_bool_t
-arm_neon_blt (pixman_implementation_t *imp,
-              uint32_t *               src_bits,
-              uint32_t *               dst_bits,
-              int                      src_stride,
-              int                      dst_stride,
-              int                      src_bpp,
-              int                      dst_bpp,
-              int                      src_x,
-              int                      src_y,
-              int                      dst_x,
-              int                      dst_y,
-              int                      width,
-              int                      height)
-{
-    if (pixman_blt_neon (
-            src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
-            src_x, src_y, dst_x, dst_y, width, height))
-    {
-	return TRUE;
-    }
-
-    return _pixman_implementation_blt (
-               imp->delegate,
-               src_bits, dst_bits, src_stride, dst_stride, src_bpp, dst_bpp,
-               src_x, src_y, dst_x, dst_y, width, height);
-}
-
-static pixman_bool_t
 arm_neon_fill (pixman_implementation_t *imp,
                uint32_t *               bits,
                int                      stride,
@@ -2765,9 +1982,6 @@ _pixman_implementation_create_arm_neon (void)
     pixman_implementation_t *imp = _pixman_implementation_create (simd);
 
     imp->composite = arm_neon_composite;
-#if 0 /* this code has some bugs */
-    imp->blt = arm_neon_blt;
-#endif
     imp->fill = arm_neon_fill;
 
     return imp;