The Tor Browser: gfx/cairo/pixman-16-bit-pipeline.patch@97036ab72558 (annotated)

gfx/cairo/pixman-16-bit-pipeline.patch@97036ab72558 (annotated)

gfx/cairo/pixman-16-bit-pipeline.patch

Tue, 06 Jan 2015 21:39:09 +0100

author: Michael Schloh von Bennewitz <michael@schloh.com>
date: Tue, 06 Jan 2015 21:39:09 +0100
branch: TOR_BUG_9701
changeset 8: 97036ab72558
permissions: -rw-r--r--

Conditionally force memory storage according to privacy.thirdparty.isolate;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.

 diff --git a/gfx/cairo/libpixman/src/pixman-access.c b/gfx/cairo/libpixman/src/pixman-access.c
 --- a/gfx/cairo/libpixman/src/pixman-access.c
 +++ b/gfx/cairo/libpixman/src/pixman-access.c
@@ -933,16 +933,54 @@ store_scanline_x2b10g10r10 (bits_image_t
      {
  	WRITE (image, pixel++,
  	       ((values[i] >> 38) & 0x3ff) |
  	       ((values[i] >> 12) & 0xffc00) |
  	       ((values[i] << 14) & 0x3ff00000));
      }
  }
 +static void
 +store_scanline_16 (bits_image_t *  image,
 +		   int             x,
 +		   int             y,
 +		   int             width,
 +		   const uint32_t *v)
 +{
 +    uint16_t *bits = (uint16_t*)(image->bits + image->rowstride * y);
 +    uint16_t *values = (uint16_t *)v;
 +    uint16_t *pixel = bits + x;
 +    int i;
 +
 +    for (i = 0; i < width; ++i)
 +    {
 +	WRITE (image, pixel++, values[i]);
 +    }
 +}
 +
 +static void
 +fetch_scanline_16 (pixman_image_t *image,
 +                            int             x,
 +                            int             y,
 +                            int             width,
 +                            uint32_t *      b,
 +                            const uint32_t *mask)
 +{
 +    const uint16_t *bits = (uint16_t*)(image->bits.bits + y * image->bits.rowstride);
 +    const uint16_t *pixel = bits + x;
 +    int i;
 +    uint16_t *buffer = (uint16_t *)b;
 +
 +    for (i = 0; i < width; ++i)
 +    {
 +	*buffer++ = READ (image, pixel++);
 +    }
 +}
 +
 +
  /*
   * Contracts a 64bpp image to 32bpp and then stores it using a regular 32-bit
   * store proc. Despite the type, this function expects a uint64_t buffer.
   */
  static void
  store_scanline_generic_64 (bits_image_t *  image,
                             int             x,
                             int             y,
@@ -1044,32 +1082,47 @@ fetch_pixel_generic_lossy_32 (bits_image
      pixman_contract (&result, &pixel64, 1);
      return result;
  }
  typedef struct
  {
      pixman_format_code_t	format;
 +    fetch_scanline_t		fetch_scanline_16;
      fetch_scanline_t		fetch_scanline_32;
      fetch_scanline_t		fetch_scanline_64;
      fetch_pixel_32_t		fetch_pixel_32;
      fetch_pixel_64_t		fetch_pixel_64;
 +    store_scanline_t		store_scanline_16;
      store_scanline_t		store_scanline_32;
      store_scanline_t		store_scanline_64;
  } format_info_t;
  #define FORMAT_INFO(format) 						\
      {									\
  	PIXMAN_ ## format,						\
 +	    NULL,							\
  	    fetch_scanline_ ## format,					\
  	    fetch_scanline_generic_64,					\
  	    fetch_pixel_ ## format, fetch_pixel_generic_64,		\
 +	    NULL,							\
  	    store_scanline_ ## format, store_scanline_generic_64	\
      }
 +#define FORMAT_INFO16(format) 						\
 +    {									\
 +	PIXMAN_ ## format,						\
 +	    fetch_scanline_16,						\
 +	    fetch_scanline_ ## format,					\
 +	    fetch_scanline_generic_64,					\
 +	    fetch_pixel_ ## format, fetch_pixel_generic_64,		\
 +	    store_scanline_16,						\
 +	    store_scanline_ ## format, store_scanline_generic_64	\
 +    }
 +
  static const format_info_t accessors[] =
  {
  /* 32 bpp formats */
      FORMAT_INFO (a8r8g8b8),
      FORMAT_INFO (x8r8g8b8),
      FORMAT_INFO (a8b8g8r8),
      FORMAT_INFO (x8b8g8r8),
@@ -1079,18 +1132,18 @@ static const format_info_t accessors[] =
      FORMAT_INFO (r8g8b8x8),
      FORMAT_INFO (x14r6g6b6),
  /* 24bpp formats */
      FORMAT_INFO (r8g8b8),
      FORMAT_INFO (b8g8r8),
  /* 16bpp formats */
 -    FORMAT_INFO (r5g6b5),
 -    FORMAT_INFO (b5g6r5),
 +    FORMAT_INFO16 (r5g6b5),
 +    FORMAT_INFO16 (b5g6r5),
      FORMAT_INFO (a1r5g5b5),
      FORMAT_INFO (x1r5g5b5),
      FORMAT_INFO (a1b5g5r5),
      FORMAT_INFO (x1b5g5r5),
      FORMAT_INFO (a4r4g4b4),
      FORMAT_INFO (x4r4g4b4),
      FORMAT_INFO (a4b4g4r4),
@@ -1132,62 +1185,64 @@ static const format_info_t accessors[] =
  /* 1bpp formats */
      FORMAT_INFO (a1),
      FORMAT_INFO (g1),
  /* Wide formats */
      { PIXMAN_a2r10g10b10,
 -      NULL, fetch_scanline_a2r10g10b10,
 +      NULL, NULL, fetch_scanline_a2r10g10b10,
        fetch_pixel_generic_lossy_32, fetch_pixel_a2r10g10b10,
        NULL, store_scanline_a2r10g10b10 },
      { PIXMAN_x2r10g10b10,
 -      NULL, fetch_scanline_x2r10g10b10,
 +      NULL, NULL, fetch_scanline_x2r10g10b10,
        fetch_pixel_generic_lossy_32, fetch_pixel_x2r10g10b10,
        NULL, store_scanline_x2r10g10b10 },
      { PIXMAN_a2b10g10r10,
 -      NULL, fetch_scanline_a2b10g10r10,
 +      NULL, NULL, fetch_scanline_a2b10g10r10,
        fetch_pixel_generic_lossy_32, fetch_pixel_a2b10g10r10,
        NULL, store_scanline_a2b10g10r10 },
      { PIXMAN_x2b10g10r10,
 -      NULL, fetch_scanline_x2b10g10r10,
 +      NULL, NULL, fetch_scanline_x2b10g10r10,
        fetch_pixel_generic_lossy_32, fetch_pixel_x2b10g10r10,
        NULL, store_scanline_x2b10g10r10 },
  /* YUV formats */
      { PIXMAN_yuy2,
 -      fetch_scanline_yuy2, fetch_scanline_generic_64,
 +      NULL, fetch_scanline_yuy2, fetch_scanline_generic_64,
        fetch_pixel_yuy2, fetch_pixel_generic_64,
        NULL, NULL },
      { PIXMAN_yv12,
 -      fetch_scanline_yv12, fetch_scanline_generic_64,
 +      NULL, fetch_scanline_yv12, fetch_scanline_generic_64,
        fetch_pixel_yv12, fetch_pixel_generic_64,
        NULL, NULL },
      { PIXMAN_null },
  };
  static void
  setup_accessors (bits_image_t *image)
  {
      const format_info_t *info = accessors;
      while (info->format != PIXMAN_null)
      {
  	if (info->format == image->format)
  	{
 +	    image->fetch_scanline_16 = info->fetch_scanline_16;
  	    image->fetch_scanline_32 = info->fetch_scanline_32;
  	    image->fetch_scanline_64 = info->fetch_scanline_64;
  	    image->fetch_pixel_32 = info->fetch_pixel_32;
  	    image->fetch_pixel_64 = info->fetch_pixel_64;
 +	    image->store_scanline_16 = info->store_scanline_16;
  	    image->store_scanline_32 = info->store_scanline_32;
  	    image->store_scanline_64 = info->store_scanline_64;
  	    return;
  	}
  	info++;
      }
 diff --git a/gfx/cairo/libpixman/src/pixman-bits-image.c b/gfx/cairo/libpixman/src/pixman-bits-image.c
 --- a/gfx/cairo/libpixman/src/pixman-bits-image.c
 +++ b/gfx/cairo/libpixman/src/pixman-bits-image.c
@@ -1247,16 +1247,31 @@ src_get_scanline_wide (pixman_iter_t *it
  void
  _pixman_bits_image_src_iter_init (pixman_image_t *image, pixman_iter_t *iter)
  {
      if (iter->flags & ITER_NARROW)
  	iter->get_scanline = src_get_scanline_narrow;
      else
  	iter->get_scanline = src_get_scanline_wide;
 +
 +}
 +
 +static uint32_t *
 +dest_get_scanline_16 (pixman_iter_t *iter, const uint32_t *mask)
 +{
 +    pixman_image_t *image  = iter->image;
 +    int             x      = iter->x;
 +    int             y      = iter->y;
 +    int             width  = iter->width;
 +    uint32_t *	    buffer = iter->buffer;
 +
 +    image->bits.fetch_scanline_16 (image, x, y, width, buffer, mask);
 +
 +    return iter->buffer;
  }
  static uint32_t *
  dest_get_scanline_narrow (pixman_iter_t *iter, const uint32_t *mask)
  {
      pixman_image_t *image  = iter->image;
      int             x      = iter->x;
      int             y      = iter->y;
@@ -1327,16 +1342,30 @@ dest_get_scanline_wide (pixman_iter_t *i
  	    free (alpha);
  	}
      }
      return iter->buffer;
  }
  static void
 +dest_write_back_16 (pixman_iter_t *iter)
 +{
 +    bits_image_t *  image  = &iter->image->bits;
 +    int             x      = iter->x;
 +    int             y      = iter->y;
 +    int             width  = iter->width;
 +    const uint32_t *buffer = iter->buffer;
 +
 +    image->store_scanline_16 (image, x, y, width, buffer);
 +
 +    iter->y++;
 +}
 +
 +static void
  dest_write_back_narrow (pixman_iter_t *iter)
  {
      bits_image_t *  image  = &iter->image->bits;
      int             x      = iter->x;
      int             y      = iter->y;
      int             width  = iter->width;
      const uint32_t *buffer = iter->buffer;
@@ -1375,28 +1404,41 @@ dest_write_back_wide (pixman_iter_t *ite
      }
      iter->y++;
  }
  void
  _pixman_bits_image_dest_iter_init (pixman_image_t *image, pixman_iter_t *iter)
  {
 -    if (iter->flags & ITER_NARROW)
 +    if (iter->flags & ITER_16)
 +    {
 +        if ((iter->flags & (ITER_IGNORE_RGB | ITER_IGNORE_ALPHA)) ==
 +	    (ITER_IGNORE_RGB | ITER_IGNORE_ALPHA))
 +	{
 +            iter->get_scanline = _pixman_iter_get_scanline_noop;
 +        }
 +        else
 +        {
 +	    iter->get_scanline = dest_get_scanline_16;
 +        }
 +	iter->write_back = dest_write_back_16;
 +    }
 +    else if (iter->flags & ITER_NARROW)
      {
  	if ((iter->flags & (ITER_IGNORE_RGB | ITER_IGNORE_ALPHA)) ==
  	    (ITER_IGNORE_RGB | ITER_IGNORE_ALPHA))
  	{
  	    iter->get_scanline = _pixman_iter_get_scanline_noop;
  	}
  	else
  	{
  	    iter->get_scanline = dest_get_scanline_narrow;
  	}
 -
 +
  	iter->write_back = dest_write_back_narrow;
      }
      else
      {
  	iter->get_scanline = dest_get_scanline_wide;
  	iter->write_back = dest_write_back_wide;
      }
  }
 diff --git a/gfx/cairo/libpixman/src/pixman-combine16.c b/gfx/cairo/libpixman/src/pixman-combine16.c
 new file mode 100644
 --- /dev/null
 +++ b/gfx/cairo/libpixman/src/pixman-combine16.c
@@ -0,0 +1,124 @@
 +#ifdef HAVE_CONFIG_H
 +#include <config.h>
 +#endif
 +
 +#include <math.h>
 +#include <string.h>
 +
 +#include "pixman-private.h"
 +
 +#include "pixman-combine32.h"
 +
 +static force_inline uint32_t
 +combine_mask (const uint32_t src, const uint32_t mask)
 +{
 +    uint32_t s, m;
 +
 +    m = mask >> A_SHIFT;
 +
 +    if (!m)
 +	return 0;
 +    s = src;
 +
 +    UN8x4_MUL_UN8 (s, m);
 +
 +    return s;
 +}
 +
 +static inline uint32_t convert_0565_to_8888(uint16_t color)
 +{
 +    return CONVERT_0565_TO_8888(color);
 +}
 +
 +static inline uint16_t convert_8888_to_0565(uint32_t color)
 +{
 +    return CONVERT_8888_TO_0565(color);
 +}
 +
 +static void
 +combine_src_u (pixman_implementation_t *imp,
 +               pixman_op_t              op,
 +               uint32_t *               dest,
 +               const uint32_t *         src,
 +               const uint32_t *         mask,
 +               int                      width)
 +{
 +    int i;
 +
 +    if (!mask)
 +	memcpy (dest, src, width * sizeof (uint16_t));
 +    else
 +    {
 +	uint16_t *d = (uint16_t*)dest;
 +	uint16_t *src16 = (uint16_t*)src;
 +	for (i = 0; i < width; ++i)
 +	{
 +	    if ((*mask & 0xff000000) == 0xff000000) {
 +		// it's likely worth special casing
 +		// fully opaque because it avoids
 +		// the cost of conversion as well the multiplication
 +		*(d + i) = *src16;
 +	    } else {
 +		// the mask is still 32bits
 +		uint32_t s = combine_mask (convert_0565_to_8888(*src16), *mask);
 +		*(d + i) = convert_8888_to_0565(s);
 +	    }
 +	    mask++;
 +	    src16++;
 +	}
 +    }
 +
 +}
 +
 +static void
 +combine_over_u (pixman_implementation_t *imp,
 +               pixman_op_t              op,
 +               uint32_t *                dest,
 +               const uint32_t *          src,
 +               const uint32_t *          mask,
 +               int                      width)
 +{
 +    int i;
 +
 +    if (!mask)
 +	memcpy (dest, src, width * sizeof (uint16_t));
 +    else
 +    {
 +	uint16_t *d = (uint16_t*)dest;
 +	uint16_t *src16 = (uint16_t*)src;
 +	for (i = 0; i < width; ++i)
 +	{
 +	    if ((*mask & 0xff000000) == 0xff000000) {
 +		// it's likely worth special casing
 +		// fully opaque because it avoids
 +		// the cost of conversion as well the multiplication
 +		*(d + i) = *src16;
 +	    } else if ((*mask & 0xff000000) == 0x00000000) {
 +		// keep the dest the same
 +	    } else {
 +		// the mask is still 32bits
 +		uint32_t s = combine_mask (convert_0565_to_8888(*src16), *mask);
 +		uint32_t ia = ALPHA_8 (~s);
 +		uint32_t d32 = convert_0565_to_8888(*(d + i));
 +		UN8x4_MUL_UN8_ADD_UN8x4 (d32, ia, s);
 +		*(d + i) = convert_8888_to_0565(d32);
 +	    }
 +	    mask++;
 +	    src16++;
 +	}
 +    }
 +
 +}
 +
 +
 +void
 +_pixman_setup_combiner_functions_16 (pixman_implementation_t *imp)
 +{
 +    int i;
 +    for (i = 0; i < PIXMAN_N_OPERATORS; i++) {
 +	imp->combine_16[i] = NULL;
 +    }
 +    imp->combine_16[PIXMAN_OP_SRC] = combine_src_u;
 +    imp->combine_16[PIXMAN_OP_OVER] = combine_over_u;
 +}
 +
 diff --git a/gfx/cairo/libpixman/src/pixman-general.c b/gfx/cairo/libpixman/src/pixman-general.c
 --- a/gfx/cairo/libpixman/src/pixman-general.c
 +++ b/gfx/cairo/libpixman/src/pixman-general.c
@@ -106,46 +106,61 @@ general_composite_rect  (pixman_implemen
      PIXMAN_COMPOSITE_ARGS (info);
      uint64_t stack_scanline_buffer[(SCANLINE_BUFFER_LENGTH * 3 + 7) / 8];
      uint8_t *scanline_buffer = (uint8_t *) stack_scanline_buffer;
      uint8_t *src_buffer, *mask_buffer, *dest_buffer;
      pixman_iter_t src_iter, mask_iter, dest_iter;
      pixman_combine_32_func_t compose;
      pixman_bool_t component_alpha;
      iter_flags_t narrow, src_flags;
 +    iter_flags_t rgb16;
      int Bpp;
      int i;
      if ((src_image->common.flags & FAST_PATH_NARROW_FORMAT)		    &&
  	(!mask_image || mask_image->common.flags & FAST_PATH_NARROW_FORMAT) &&
  	(dest_image->common.flags & FAST_PATH_NARROW_FORMAT))
      {
  	narrow = ITER_NARROW;
  	Bpp = 4;
      }
      else
      {
  	narrow = 0;
  	Bpp = 8;
      }
 +    // XXX: This special casing is bad. Ideally, we'd keep the general code general perhaps
 +    // by having it deal more specifically with different intermediate formats
 +    if (
 +	(dest_image->common.flags & FAST_PATH_16_FORMAT && (src_image->type == LINEAR || src_image->type == RADIAL)) &&
 +	( op == PIXMAN_OP_SRC ||
 +         (op == PIXMAN_OP_OVER && (src_image->common.flags & FAST_PATH_IS_OPAQUE))
 +	)
 +	) {
 +	rgb16 = ITER_16;
 +    } else {
 +	rgb16 = 0;
 +    }
 +
 +
      if (width * Bpp > SCANLINE_BUFFER_LENGTH)
      {
  	scanline_buffer = pixman_malloc_abc (width, 3, Bpp);
  	if (!scanline_buffer)
  	    return;
      }
      src_buffer = scanline_buffer;
      mask_buffer = src_buffer + width * Bpp;
      dest_buffer = mask_buffer + width * Bpp;
      /* src iter */
 -    src_flags = narrow | op_flags[op].src;
 +    src_flags = narrow | op_flags[op].src | rgb16;
      _pixman_implementation_src_iter_init (imp->toplevel, &src_iter, src_image,
  					  src_x, src_y, width, height,
  					  src_buffer, src_flags);
      /* mask iter */
      if ((src_flags & (ITER_IGNORE_ALPHA | ITER_IGNORE_RGB)) ==
  	(ITER_IGNORE_ALPHA | ITER_IGNORE_RGB))
@@ -164,20 +179,20 @@ general_composite_rect  (pixman_implemen
      _pixman_implementation_src_iter_init (
  	imp->toplevel, &mask_iter, mask_image, mask_x, mask_y, width, height,
  	mask_buffer, narrow | (component_alpha? 0 : ITER_IGNORE_RGB));
      /* dest iter */
      _pixman_implementation_dest_iter_init (
  	imp->toplevel, &dest_iter, dest_image, dest_x, dest_y, width, height,
 -	dest_buffer, narrow | op_flags[op].dst);
 +	dest_buffer, narrow | op_flags[op].dst | rgb16);
      compose = _pixman_implementation_lookup_combiner (
 -	imp->toplevel, op, component_alpha, narrow);
 +	imp->toplevel, op, component_alpha, narrow, !!rgb16);
      if (!compose)
  	return;
      for (i = 0; i < height; ++i)
      {
  	uint32_t *s, *m, *d;
@@ -234,16 +249,17 @@ general_fill (pixman_implementation_t *i
      return FALSE;
  }
  pixman_implementation_t *
  _pixman_implementation_create_general (void)
  {
      pixman_implementation_t *imp = _pixman_implementation_create (NULL, general_fast_path);
 +    _pixman_setup_combiner_functions_16 (imp);
      _pixman_setup_combiner_functions_32 (imp);
      _pixman_setup_combiner_functions_64 (imp);
      imp->blt = general_blt;
      imp->fill = general_fill;
      imp->src_iter_init = general_src_iter_init;
      imp->dest_iter_init = general_dest_iter_init;
 diff --git a/gfx/cairo/libpixman/src/pixman-image.c b/gfx/cairo/libpixman/src/pixman-image.c
 --- a/gfx/cairo/libpixman/src/pixman-image.c
 +++ b/gfx/cairo/libpixman/src/pixman-image.c
@@ -451,16 +451,20 @@ compute_image_info (pixman_image_t *imag
  		flags |= FAST_PATH_IS_OPAQUE;
  	}
  	if (image->bits.read_func || image->bits.write_func)
  	    flags &= ~FAST_PATH_NO_ACCESSORS;
  	if (PIXMAN_FORMAT_IS_WIDE (image->bits.format))
  	    flags &= ~FAST_PATH_NARROW_FORMAT;
 +
 +	if (image->bits.format == PIXMAN_r5g6b5)
 +	    flags |= FAST_PATH_16_FORMAT;
 +
  	break;
      case RADIAL:
  	code = PIXMAN_unknown;
  	/*
  	 * As explained in pixman-radial-gradient.c, every point of
  	 * the plane has a valid associated radius (and thus will be
 diff --git a/gfx/cairo/libpixman/src/pixman-implementation.c b/gfx/cairo/libpixman/src/pixman-implementation.c
 --- a/gfx/cairo/libpixman/src/pixman-implementation.c
 +++ b/gfx/cairo/libpixman/src/pixman-implementation.c
@@ -101,45 +101,51 @@ pixman_implementation_t *
      imp->fill = delegate_fill;
      imp->src_iter_init = delegate_src_iter_init;
      imp->dest_iter_init = delegate_dest_iter_init;
      imp->fast_paths = fast_paths;
      for (i = 0; i < PIXMAN_N_OPERATORS; ++i)
      {
 +	imp->combine_16[i] = NULL;
  	imp->combine_32[i] = NULL;
  	imp->combine_64[i] = NULL;
  	imp->combine_32_ca[i] = NULL;
  	imp->combine_64_ca[i] = NULL;
      }
      return imp;
  }
  pixman_combine_32_func_t
  _pixman_implementation_lookup_combiner (pixman_implementation_t *imp,
  					pixman_op_t		 op,
  					pixman_bool_t		 component_alpha,
 -					pixman_bool_t		 narrow)
 +					pixman_bool_t		 narrow,
 +					pixman_bool_t		 rgb16)
  {
      pixman_combine_32_func_t f;
      do
      {
  	pixman_combine_32_func_t (*combiners[]) =
  	{
  	    (pixman_combine_32_func_t *)imp->combine_64,
  	    (pixman_combine_32_func_t *)imp->combine_64_ca,
  	    imp->combine_32,
  	    imp->combine_32_ca,
 +	    (pixman_combine_32_func_t *)imp->combine_16,
 +	    NULL,
  	};
 -
 -	f = combiners[component_alpha | (narrow << 1)][op];
 -
 +        if (rgb16) {
 +            f = combiners[4][op];
 +        } else {
 +            f = combiners[component_alpha + (narrow << 1)][op];
 +        }
  	imp = imp->delegate;
      }
      while (!f);
      return f;
  }
  pixman_bool_t
 diff --git a/gfx/cairo/libpixman/src/pixman-linear-gradient.c b/gfx/cairo/libpixman/src/pixman-linear-gradient.c
 --- a/gfx/cairo/libpixman/src/pixman-linear-gradient.c
 +++ b/gfx/cairo/libpixman/src/pixman-linear-gradient.c
@@ -217,42 +217,185 @@ linear_get_scanline_narrow (pixman_iter_
  	}
      }
      iter->y++;
      return iter->buffer;
  }
 +static uint16_t convert_8888_to_0565(uint32_t color)
 +{
 +    return CONVERT_8888_TO_0565(color);
 +}
 +
 +static uint32_t *
 +linear_get_scanline_16 (pixman_iter_t  *iter,
 +			const uint32_t *mask)
 +{
 +    pixman_image_t *image  = iter->image;
 +    int             x      = iter->x;
 +    int             y      = iter->y;
 +    int             width  = iter->width;
 +    uint16_t *      buffer = (uint16_t*)iter->buffer;
 +
 +    pixman_vector_t v, unit;
 +    pixman_fixed_32_32_t l;
 +    pixman_fixed_48_16_t dx, dy;
 +    gradient_t *gradient = (gradient_t *)image;
 +    linear_gradient_t *linear = (linear_gradient_t *)image;
 +    uint16_t *end = buffer + width;
 +    pixman_gradient_walker_t walker;
 +
 +    _pixman_gradient_walker_init (&walker, gradient, image->common.repeat);
 +
 +    /* reference point is the center of the pixel */
 +    v.vector[0] = pixman_int_to_fixed (x) + pixman_fixed_1 / 2;
 +    v.vector[1] = pixman_int_to_fixed (y) + pixman_fixed_1 / 2;
 +    v.vector[2] = pixman_fixed_1;
 +
 +    if (image->common.transform)
 +    {
 +	if (!pixman_transform_point_3d (image->common.transform, &v))
 +	    return iter->buffer;
 +
 +	unit.vector[0] = image->common.transform->matrix[0][0];
 +	unit.vector[1] = image->common.transform->matrix[1][0];
 +	unit.vector[2] = image->common.transform->matrix[2][0];
 +    }
 +    else
 +    {
 +	unit.vector[0] = pixman_fixed_1;
 +	unit.vector[1] = 0;
 +	unit.vector[2] = 0;
 +    }
 +
 +    dx = linear->p2.x - linear->p1.x;
 +    dy = linear->p2.y - linear->p1.y;
 +
 +    l = dx * dx + dy * dy;
 +
 +    if (l == 0 || unit.vector[2] == 0)
 +    {
 +	/* affine transformation only */
 +        pixman_fixed_32_32_t t, next_inc;
 +	double inc;
 +
 +	if (l == 0 || v.vector[2] == 0)
 +	{
 +	    t = 0;
 +	    inc = 0;
 +	}
 +	else
 +	{
 +	    double invden, v2;
 +
 +	    invden = pixman_fixed_1 * (double) pixman_fixed_1 /
 +		(l * (double) v.vector[2]);
 +	    v2 = v.vector[2] * (1. / pixman_fixed_1);
 +	    t = ((dx * v.vector[0] + dy * v.vector[1]) -
 +		 (dx * linear->p1.x + dy * linear->p1.y) * v2) * invden;
 +	    inc = (dx * unit.vector[0] + dy * unit.vector[1]) * invden;
 +	}
 +	next_inc = 0;
 +
 +	if (((pixman_fixed_32_32_t )(inc * width)) == 0)
 +	{
 +	    register uint16_t color;
 +
 +	    color = convert_8888_to_0565(_pixman_gradient_walker_pixel (&walker, t));
 +	    while (buffer < end)
 +		*buffer++ = color;
 +	}
 +	else
 +	{
 +	    int i;
 +
 +	    i = 0;
 +	    while (buffer < end)
 +	    {
 +		if (!mask || *mask++)
 +		{
 +		    *buffer = convert_8888_to_0565(_pixman_gradient_walker_pixel (&walker,
 +										  t + next_inc));
 +		}
 +		i++;
 +		next_inc = inc * i;
 +		buffer++;
 +	    }
 +	}
 +    }
 +    else
 +    {
 +	/* projective transformation */
 +        double t;
 +
 +	t = 0;
 +
 +	while (buffer < end)
 +	{
 +	    if (!mask || *mask++)
 +	    {
 +	        if (v.vector[2] != 0)
 +		{
 +		    double invden, v2;
 +
 +		    invden = pixman_fixed_1 * (double) pixman_fixed_1 /
 +			(l * (double) v.vector[2]);
 +		    v2 = v.vector[2] * (1. / pixman_fixed_1);
 +		    t = ((dx * v.vector[0] + dy * v.vector[1]) -
 +			 (dx * linear->p1.x + dy * linear->p1.y) * v2) * invden;
 +		}
 +
 +		*buffer = convert_8888_to_0565(_pixman_gradient_walker_pixel (&walker, t));
 +	    }
 +
 +	    ++buffer;
 +
 +	    v.vector[0] += unit.vector[0];
 +	    v.vector[1] += unit.vector[1];
 +	    v.vector[2] += unit.vector[2];
 +	}
 +    }
 +
 +    iter->y++;
 +
 +    return iter->buffer;
 +}
 +
  static uint32_t *
  linear_get_scanline_wide (pixman_iter_t *iter, const uint32_t *mask)
  {
      uint32_t *buffer = linear_get_scanline_narrow (iter, NULL);
      pixman_expand ((uint64_t *)buffer, buffer, PIXMAN_a8r8g8b8, iter->width);
      return buffer;
  }
  void
  _pixman_linear_gradient_iter_init (pixman_image_t *image, pixman_iter_t  *iter)
  {
      if (linear_gradient_is_horizontal (
  	    iter->image, iter->x, iter->y, iter->width, iter->height))
      {
 -	if (iter->flags & ITER_NARROW)
 +	if (iter->flags & ITER_16)
 +	    linear_get_scanline_16 (iter, NULL);
 +	else if (iter->flags & ITER_NARROW)
  	    linear_get_scanline_narrow (iter, NULL);
  	else
  	    linear_get_scanline_wide (iter, NULL);
  	iter->get_scanline = _pixman_iter_get_scanline_noop;
      }
      else
      {
 -	if (iter->flags & ITER_NARROW)
 +	if (iter->flags & ITER_16)
 +	    iter->get_scanline = linear_get_scanline_16;
 +	else if (iter->flags & ITER_NARROW)
  	    iter->get_scanline = linear_get_scanline_narrow;
  	else
  	    iter->get_scanline = linear_get_scanline_wide;
      }
  }
  PIXMAN_EXPORT pixman_image_t *
  pixman_image_create_linear_gradient (pixman_point_fixed_t *        p1,
 diff --git a/gfx/cairo/libpixman/src/pixman-private.h b/gfx/cairo/libpixman/src/pixman-private.h
 --- a/gfx/cairo/libpixman/src/pixman-private.h
 +++ b/gfx/cairo/libpixman/src/pixman-private.h
@@ -152,24 +152,28 @@ struct bits_image
      int                        height;
      uint32_t *                 bits;
      uint32_t *                 free_me;
      int                        rowstride;  /* in number of uint32_t's */
      fetch_scanline_t           get_scanline_32;
      fetch_scanline_t           get_scanline_64;
 +    fetch_scanline_t           fetch_scanline_16;
 +
      fetch_scanline_t           fetch_scanline_32;
      fetch_pixel_32_t	       fetch_pixel_32;
      store_scanline_t           store_scanline_32;
      fetch_scanline_t           fetch_scanline_64;
      fetch_pixel_64_t	       fetch_pixel_64;
      store_scanline_t           store_scanline_64;
 +    store_scanline_t           store_scanline_16;
 +
      /* Used for indirect access to the bits */
      pixman_read_memory_func_t  read_func;
      pixman_write_memory_func_t write_func;
  };
  union pixman_image
  {
      image_type_t       type;
@@ -202,17 +206,24 @@ typedef enum
       * destination.
       *
       * When he destination is xRGB, this is useful knowledge, because then
       * we can treat it as if it were ARGB, which means in some cases we can
       * avoid copying it to a temporary buffer.
       */
      ITER_LOCALIZED_ALPHA =	(1 << 1),
      ITER_IGNORE_ALPHA =		(1 << 2),
 -    ITER_IGNORE_RGB =		(1 << 3)
 +    ITER_IGNORE_RGB =		(1 << 3),
 +
 +    /* With the addition of ITER_16 we now have two flags that to represent
 +     * 3 pipelines. This means that there can be an invalid state when
 +     * both ITER_NARROW and ITER_16 are set. In this case
 +     * ITER_16 overrides NARROW and we should use the 16 bit pipeline.
 +     * Note: ITER_16 still has a 32 bit mask, which is a bit weird. */
 +    ITER_16 =			(1 << 4)
  } iter_flags_t;
  struct pixman_iter_t
  {
      /* These are initialized by _pixman_implementation_{src,dest}_init */
      pixman_image_t *		image;
      uint32_t *			buffer;
      int				x, y;
@@ -429,16 +440,17 @@ typedef pixman_bool_t (*pixman_fill_func
  					     int                      x,
  					     int                      y,
  					     int                      width,
  					     int                      height,
  					     uint32_t                 xor);
  typedef void (*pixman_iter_init_func_t) (pixman_implementation_t *imp,
                                           pixman_iter_t           *iter);
 +void _pixman_setup_combiner_functions_16 (pixman_implementation_t *imp);
  void _pixman_setup_combiner_functions_32 (pixman_implementation_t *imp);
  void _pixman_setup_combiner_functions_64 (pixman_implementation_t *imp);
  typedef struct
  {
      pixman_op_t             op;
      pixman_format_code_t    src_format;
      uint32_t		    src_flags;
@@ -459,32 +471,34 @@ struct pixman_implementation_t
      pixman_fill_func_t		fill;
      pixman_iter_init_func_t     src_iter_init;
      pixman_iter_init_func_t     dest_iter_init;
      pixman_combine_32_func_t	combine_32[PIXMAN_N_OPERATORS];
      pixman_combine_32_func_t	combine_32_ca[PIXMAN_N_OPERATORS];
      pixman_combine_64_func_t	combine_64[PIXMAN_N_OPERATORS];
      pixman_combine_64_func_t	combine_64_ca[PIXMAN_N_OPERATORS];
 +    pixman_combine_64_func_t	combine_16[PIXMAN_N_OPERATORS];
  };
  uint32_t
  _pixman_image_get_solid (pixman_implementation_t *imp,
  			 pixman_image_t *         image,
                           pixman_format_code_t     format);
  pixman_implementation_t *
  _pixman_implementation_create (pixman_implementation_t *delegate,
  			       const pixman_fast_path_t *fast_paths);
  pixman_combine_32_func_t
  _pixman_implementation_lookup_combiner (pixman_implementation_t *imp,
  					pixman_op_t		 op,
  					pixman_bool_t		 component_alpha,
 -					pixman_bool_t		 wide);
 +					pixman_bool_t		 wide,
 +					pixman_bool_t		 rgb16);
  pixman_bool_t
  _pixman_implementation_blt (pixman_implementation_t *imp,
                              uint32_t *               src_bits,
                              uint32_t *               dst_bits,
                              int                      src_stride,
                              int                      dst_stride,
                              int                      src_bpp,
@@ -613,16 +627,17 @@ uint32_t *
  #define FAST_PATH_Y_UNIT_ZERO			(1 << 18)
  #define FAST_PATH_BILINEAR_FILTER		(1 << 19)
  #define FAST_PATH_ROTATE_90_TRANSFORM		(1 << 20)
  #define FAST_PATH_ROTATE_180_TRANSFORM		(1 << 21)
  #define FAST_PATH_ROTATE_270_TRANSFORM		(1 << 22)
  #define FAST_PATH_SAMPLES_COVER_CLIP_NEAREST	(1 << 23)
  #define FAST_PATH_SAMPLES_COVER_CLIP_BILINEAR	(1 << 24)
  #define FAST_PATH_BITS_IMAGE			(1 << 25)
 +#define FAST_PATH_16_FORMAT			(1 << 26)
  #define FAST_PATH_PAD_REPEAT						\
      (FAST_PATH_NO_NONE_REPEAT		|				\
       FAST_PATH_NO_NORMAL_REPEAT		|				\
       FAST_PATH_NO_REFLECT_REPEAT)
  #define FAST_PATH_NORMAL_REPEAT						\
      (FAST_PATH_NO_NONE_REPEAT		|				\
 diff --git a/gfx/cairo/libpixman/src/pixman-radial-gradient.c b/gfx/cairo/libpixman/src/pixman-radial-gradient.c
 --- a/gfx/cairo/libpixman/src/pixman-radial-gradient.c
 +++ b/gfx/cairo/libpixman/src/pixman-radial-gradient.c
@@ -395,35 +395,289 @@ radial_get_scanline_narrow (pixman_iter_
  	    v.vector[2] += unit.vector[2];
  	}
      }
      iter->y++;
      return iter->buffer;
  }
 +static uint16_t convert_8888_to_0565(uint32_t color)
 +{
 +    return CONVERT_8888_TO_0565(color);
 +}
 +
 +static uint32_t *
 +radial_get_scanline_16 (pixman_iter_t *iter, const uint32_t *mask)
 +{
 +    /*
 +     * Implementation of radial gradients following the PDF specification.
 +     * See section 8.7.4.5.4 Type 3 (Radial) Shadings of the PDF Reference
 +     * Manual (PDF 32000-1:2008 at the time of this writing).
 +     *
 +     * In the radial gradient problem we are given two circles (c₁,r₁) and
 +     * (c₂,r₂) that define the gradient itself.
 +     *
 +     * Mathematically the gradient can be defined as the family of circles
 +     *
 +     *     ((1-t)·c₁ + t·(c₂), (1-t)·r₁ + t·r₂)
 +     *
 +     * excluding those circles whose radius would be < 0. When a point
 +     * belongs to more than one circle, the one with a bigger t is the only
 +     * one that contributes to its color. When a point does not belong
 +     * to any of the circles, it is transparent black, i.e. RGBA (0, 0, 0, 0).
 +     * Further limitations on the range of values for t are imposed when
 +     * the gradient is not repeated, namely t must belong to [0,1].
 +     *
 +     * The graphical result is the same as drawing the valid (radius > 0)
 +     * circles with increasing t in [-inf, +inf] (or in [0,1] if the gradient
 +     * is not repeated) using SOURCE operator composition.
 +     *
 +     * It looks like a cone pointing towards the viewer if the ending circle
 +     * is smaller than the starting one, a cone pointing inside the page if
 +     * the starting circle is the smaller one and like a cylinder if they
 +     * have the same radius.
 +     *
 +     * What we actually do is, given the point whose color we are interested
 +     * in, compute the t values for that point, solving for t in:
 +     *
 +     *     length((1-t)·c₁ + t·(c₂) - p) = (1-t)·r₁ + t·r₂
 +     *
 +     * Let's rewrite it in a simpler way, by defining some auxiliary
 +     * variables:
 +     *
 +     *     cd = c₂ - c₁
 +     *     pd = p - c₁
 +     *     dr = r₂ - r₁
 +     *     length(t·cd - pd) = r₁ + t·dr
 +     *
 +     * which actually means
 +     *
 +     *     hypot(t·cdx - pdx, t·cdy - pdy) = r₁ + t·dr
 +     *
 +     * or
 +     *
 +     *     ⎷((t·cdx - pdx)² + (t·cdy - pdy)²) = r₁ + t·dr.
 +     *
 +     * If we impose (as stated earlier) that r₁ + t·dr >= 0, it becomes:
 +     *
 +     *     (t·cdx - pdx)² + (t·cdy - pdy)² = (r₁ + t·dr)²
 +     *
 +     * where we can actually expand the squares and solve for t:
 +     *
 +     *     t²cdx² - 2t·cdx·pdx + pdx² + t²cdy² - 2t·cdy·pdy + pdy² =
 +     *       = r₁² + 2·r₁·t·dr + t²·dr²
 +     *
 +     *     (cdx² + cdy² - dr²)t² - 2(cdx·pdx + cdy·pdy + r₁·dr)t +
 +     *         (pdx² + pdy² - r₁²) = 0
 +     *
 +     *     A = cdx² + cdy² - dr²
 +     *     B = pdx·cdx + pdy·cdy + r₁·dr
 +     *     C = pdx² + pdy² - r₁²
 +     *     At² - 2Bt + C = 0
 +     *
 +     * The solutions (unless the equation degenerates because of A = 0) are:
 +     *
 +     *     t = (B ± ⎷(B² - A·C)) / A
 +     *
 +     * The solution we are going to prefer is the bigger one, unless the
 +     * radius associated to it is negative (or it falls outside the valid t
 +     * range).
 +     *
 +     * Additional observations (useful for optimizations):
 +     * A does not depend on p
 +     *
 +     * A < 0 <=> one of the two circles completely contains the other one
 +     *   <=> for every p, the radiuses associated with the two t solutions
 +     *       have opposite sign
 +     */
 +    pixman_image_t *image = iter->image;
 +    int x = iter->x;
 +    int y = iter->y;
 +    int width = iter->width;
 +    uint16_t *buffer = iter->buffer;
 +
 +    gradient_t *gradient = (gradient_t *)image;
 +    radial_gradient_t *radial = (radial_gradient_t *)image;
 +    uint16_t *end = buffer + width;
 +    pixman_gradient_walker_t walker;
 +    pixman_vector_t v, unit;
 +
 +    /* reference point is the center of the pixel */
 +    v.vector[0] = pixman_int_to_fixed (x) + pixman_fixed_1 / 2;
 +    v.vector[1] = pixman_int_to_fixed (y) + pixman_fixed_1 / 2;
 +    v.vector[2] = pixman_fixed_1;
 +
 +    _pixman_gradient_walker_init (&walker, gradient, image->common.repeat);
 +
 +    if (image->common.transform)
 +    {
 +	if (!pixman_transform_point_3d (image->common.transform, &v))
 +	    return iter->buffer;
 +
 +	unit.vector[0] = image->common.transform->matrix[0][0];
 +	unit.vector[1] = image->common.transform->matrix[1][0];
 +	unit.vector[2] = image->common.transform->matrix[2][0];
 +    }
 +    else
 +    {
 +	unit.vector[0] = pixman_fixed_1;
 +	unit.vector[1] = 0;
 +	unit.vector[2] = 0;
 +    }
 +
 +    if (unit.vector[2] == 0 && v.vector[2] == pixman_fixed_1)
 +    {
 +	/*
 +	 * Given:
 +	 *
 +	 * t = (B ± ⎷(B² - A·C)) / A
 +	 *
 +	 * where
 +	 *
 +	 * A = cdx² + cdy² - dr²
 +	 * B = pdx·cdx + pdy·cdy + r₁·dr
 +	 * C = pdx² + pdy² - r₁²
 +	 * det = B² - A·C
 +	 *
 +	 * Since we have an affine transformation, we know that (pdx, pdy)
 +	 * increase linearly with each pixel,
 +	 *
 +	 * pdx = pdx₀ + n·ux,
 +	 * pdy = pdy₀ + n·uy,
 +	 *
 +	 * we can then express B, C and det through multiple differentiation.
 +	 */
 +	pixman_fixed_32_32_t b, db, c, dc, ddc;
 +
 +	/* warning: this computation may overflow */
 +	v.vector[0] -= radial->c1.x;
 +	v.vector[1] -= radial->c1.y;
 +
 +	/*
 +	 * B and C are computed and updated exactly.
 +	 * If fdot was used instead of dot, in the worst case it would
 +	 * lose 11 bits of precision in each of the multiplication and
 +	 * summing up would zero out all the bit that were preserved,
 +	 * thus making the result 0 instead of the correct one.
 +	 * This would mean a worst case of unbound relative error or
 +	 * about 2^10 absolute error
 +	 */
 +	b = dot (v.vector[0], v.vector[1], radial->c1.radius,
 +		 radial->delta.x, radial->delta.y, radial->delta.radius);
 +	db = dot (unit.vector[0], unit.vector[1], 0,
 +		  radial->delta.x, radial->delta.y, 0);
 +
 +	c = dot (v.vector[0], v.vector[1],
 +		 -((pixman_fixed_48_16_t) radial->c1.radius),
 +		 v.vector[0], v.vector[1], radial->c1.radius);
 +	dc = dot (2 * (pixman_fixed_48_16_t) v.vector[0] + unit.vector[0],
 +		  2 * (pixman_fixed_48_16_t) v.vector[1] + unit.vector[1],
 +		  0,
 +		  unit.vector[0], unit.vector[1], 0);
 +	ddc = 2 * dot (unit.vector[0], unit.vector[1], 0,
 +		       unit.vector[0], unit.vector[1], 0);
 +
 +	while (buffer < end)
 +	{
 +	    if (!mask || *mask++)
 +	    {
 +		*buffer = convert_8888_to_0565(
 +			  radial_compute_color (radial->a, b, c,
 +						radial->inva,
 +						radial->delta.radius,
 +						radial->mindr,
 +						&walker,
 +						image->common.repeat));
 +	    }
 +
 +	    b += db;
 +	    c += dc;
 +	    dc += ddc;
 +	    ++buffer;
 +	}
 +    }
 +    else
 +    {
 +	/* projective */
 +	/* Warning:
 +	 * error propagation guarantees are much looser than in the affine case
 +	 */
 +	while (buffer < end)
 +	{
 +	    if (!mask || *mask++)
 +	    {
 +		if (v.vector[2] != 0)
 +		{
 +		    double pdx, pdy, invv2, b, c;
 +
 +		    invv2 = 1. * pixman_fixed_1 / v.vector[2];
 +
 +		    pdx = v.vector[0] * invv2 - radial->c1.x;
 +		    /*    / pixman_fixed_1 */
 +
 +		    pdy = v.vector[1] * invv2 - radial->c1.y;
 +		    /*    / pixman_fixed_1 */
 +
 +		    b = fdot (pdx, pdy, radial->c1.radius,
 +			      radial->delta.x, radial->delta.y,
 +			      radial->delta.radius);
 +		    /*  / pixman_fixed_1 / pixman_fixed_1 */
 +
 +		    c = fdot (pdx, pdy, -radial->c1.radius,
 +			      pdx, pdy, radial->c1.radius);
 +		    /*  / pixman_fixed_1 / pixman_fixed_1 */
 +
 +		    *buffer = convert_8888_to_0565 (
 +			      radial_compute_color (radial->a, b, c,
 +						    radial->inva,
 +						    radial->delta.radius,
 +						    radial->mindr,
 +						    &walker,
 +						    image->common.repeat));
 +		}
 +		else
 +		{
 +		    *buffer = 0;
 +		}
 +	    }
 +
 +	    ++buffer;
 +
 +	    v.vector[0] += unit.vector[0];
 +	    v.vector[1] += unit.vector[1];
 +	    v.vector[2] += unit.vector[2];
 +	}
 +    }
 +
 +    iter->y++;
 +    return iter->buffer;
 +}
  static uint32_t *
  radial_get_scanline_wide (pixman_iter_t *iter, const uint32_t *mask)
  {
      uint32_t *buffer = radial_get_scanline_narrow (iter, NULL);
      pixman_expand ((uint64_t *)buffer, buffer, PIXMAN_a8r8g8b8, iter->width);
      return buffer;
  }
  void
  _pixman_radial_gradient_iter_init (pixman_image_t *image, pixman_iter_t *iter)
  {
 -    if (iter->flags & ITER_NARROW)
 +    if (iter->flags & ITER_16)
 +	iter->get_scanline = radial_get_scanline_16;
 +    else if (iter->flags & ITER_NARROW)
  	iter->get_scanline = radial_get_scanline_narrow;
      else
  	iter->get_scanline = radial_get_scanline_wide;
  }
 +
  PIXMAN_EXPORT pixman_image_t *
  pixman_image_create_radial_gradient (pixman_point_fixed_t *        inner,
                                       pixman_point_fixed_t *        outer,
                                       pixman_fixed_t                inner_radius,
                                       pixman_fixed_t                outer_radius,
                                       const pixman_gradient_stop_t *stops,
                                       int                           n_stops)
  {

The Tor Browser / annotate

gfx/cairo/pixman-16-bit-pipeline.patch@97036ab72558 (annotated)

gfx/cairo/pixman-16-bit-pipeline.patch