1 /*

     2  * Copyright © 2011,2012  Google, Inc.

     3  *

     4  *  This is part of HarfBuzz, a text shaping library.

     5  *

     6  * Permission is hereby granted, without written agreement and without

     7  * license or royalty fees, to use, copy, modify, and distribute this

     8  * software and its documentation for any purpose, provided that the

     9  * above copyright notice and the following two paragraphs appear in

    10  * all copies of this software.

    11  *

    12  * IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE TO ANY PARTY FOR

    13  * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES

    14  * ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN

    15  * IF THE COPYRIGHT HOLDER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH

    16  * DAMAGE.

    17  *

    18  * THE COPYRIGHT HOLDER SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING,

    19  * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND

    20  * FITNESS FOR A PARTICULAR PURPOSE.  THE SOFTWARE PROVIDED HEREUNDER IS

    21  * ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO

    22  * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.

    23  *

    24  * Google Author(s): Behdad Esfahbod

    25  */

    26 

    27 #include "hb-ot-shape-normalize-private.hh"

    28 #include "hb-ot-shape-complex-private.hh"

    29 #include "hb-ot-shape-private.hh"

    30 

    31 

    32 /*

    33  * HIGHLEVEL DESIGN:

    34  *

    35  * This file exports one main function: _hb_ot_shape_normalize().

    36  *

    37  * This function closely reflects the Unicode Normalization Algorithm,

    38  * yet it's different.

    39  *

    40  * Each shaper specifies whether it prefers decomposed (NFD) or composed (NFC).

    41  * The logic however tries to use whatever the font can support.

    42  *

    43  * In general what happens is that: each grapheme is decomposed in a chain

    44  * of 1:2 decompositions, marks reordered, and then recomposed if desired,

    45  * so far it's like Unicode Normalization.  However, the decomposition and

    46  * recomposition only happens if the font supports the resulting characters.

    47  *

    48  * The goals are:

    49  *

    50  *   - Try to render all canonically equivalent strings similarly.  To really

    51  *     achieve this we have to always do the full decomposition and then

    52  *     selectively recompose from there.  It's kinda too expensive though, so

    53  *     we skip some cases.  For example, if composed is desired, we simply

    54  *     don't touch 1-character clusters that are supported by the font, even

    55  *     though their NFC may be different.

    56  *

    57  *   - When a font has a precomposed character for a sequence but the 'ccmp'

    58  *     feature in the font is not adequate, use the precomposed character

    59  *     which typically has better mark positioning.

    60  *

    61  *   - When a font does not support a combining mark, but supports it precomposed

    62  *     with previous base, use that.  This needs the itemizer to have this

    63  *     knowledge too.  We need to provide assistance to the itemizer.

    64  *

    65  *   - When a font does not support a character but supports its decomposition,

    66  *     well, use the decomposition (preferring the canonical decomposition, but

    67  *     falling back to the compatibility decomposition if necessary).  The

    68  *     compatibility decomposition is really nice to have, for characters like

    69  *     ellipsis, or various-sized space characters.

    70  *

    71  *   - The complex shapers can customize the compose and decompose functions to

    72  *     offload some of their requirements to the normalizer.  For example, the

    73  *     Indic shaper may want to disallow recomposing of two matras.

    74  *

    75  *   - We try compatibility decomposition if decomposing through canonical

    76  *     decomposition alone failed to find a sequence that the font supports.

    77  *     We don't try compatibility decomposition recursively during the canonical

    78  *     decomposition phase.  This has minimal impact.  There are only a handful

    79  *     of Greek letter that have canonical decompositions that include characters

    80  *     with compatibility decomposition.  Those can be found using this command:

    81  *

    82  *     egrep  "`echo -n ';('; grep ';<' UnicodeData.txt | cut -d';' -f1 | tr '\n' '|'; echo ') '`" UnicodeData.txt

    83  */

    84 

    85 static bool

    86 decompose_unicode (const hb_ot_shape_normalize_context_t *c,

    87                    hb_codepoint_t  ab,

    88                    hb_codepoint_t *a,

    89                    hb_codepoint_t *b)

    90 {

    91   return c->unicode->decompose (ab, a, b);

    92 }

    93 

    94 static bool

    95 compose_unicode (const hb_ot_shape_normalize_context_t *c,

    96                  hb_codepoint_t  a,

    97                  hb_codepoint_t  b,

    98                  hb_codepoint_t *ab)

    99 {

   100   return c->unicode->compose (a, b, ab);

   101 }

   102 

   103 static inline void

   104 set_glyph (hb_glyph_info_t &info, hb_font_t *font)

   105 {

   106   font->get_glyph (info.codepoint, 0, &info.glyph_index());

   107 }

   108 

   109 static inline void

   110 output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph)

   111 {

   112   buffer->cur().glyph_index() = glyph;

   113   buffer->output_glyph (unichar);

   114   _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer->unicode);

   115 }

   116 

   117 static inline void

   118 next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)

   119 {

   120   buffer->cur().glyph_index() = glyph;

   121   buffer->next_glyph ();

   122 }

   123 

   124 static inline void

   125 skip_char (hb_buffer_t *buffer)

   126 {

   127   buffer->skip_glyph ();

   128 }

   129 

   130 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */

   131 static inline unsigned int

   132 decompose (const hb_ot_shape_normalize_context_t *c, bool shortest, hb_codepoint_t ab)

   133 {

   134   hb_codepoint_t a, b, a_glyph, b_glyph;

   135   hb_buffer_t * const buffer = c->buffer;

   136   hb_font_t * const font = c->font;

   137 

   138   if (!c->decompose (c, ab, &a, &b) ||

   139       (b && !font->get_glyph (b, 0, &b_glyph)))

   140     return 0;

   141 

   142   bool has_a = font->get_glyph (a, 0, &a_glyph);

   143   if (shortest && has_a) {

   144     /* Output a and b */

   145     output_char (buffer, a, a_glyph);

   146     if (likely (b)) {

   147       output_char (buffer, b, b_glyph);

   148       return 2;

   149     }

   150     return 1;

   151   }

   152 

   153   unsigned int ret;

   154   if ((ret = decompose (c, shortest, a))) {

   155     if (b) {

   156       output_char (buffer, b, b_glyph);

   157       return ret + 1;

   158     }

   159     return ret;

   160   }

   161 

   162   if (has_a) {

   163     output_char (buffer, a, a_glyph);

   164     if (likely (b)) {

   165       output_char (buffer, b, b_glyph);

   166       return 2;

   167     }

   168     return 1;

   169   }

   170 

   171   return 0;

   172 }

   173 

   174 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */

   175 static inline unsigned int

   176 decompose_compatibility (const hb_ot_shape_normalize_context_t *c, hb_codepoint_t u)

   177 {

   178   unsigned int len, i;

   179   hb_codepoint_t decomposed[HB_UNICODE_MAX_DECOMPOSITION_LEN];

   180   hb_codepoint_t glyphs[HB_UNICODE_MAX_DECOMPOSITION_LEN];

   181 

   182   len = c->buffer->unicode->decompose_compatibility (u, decomposed);

   183   if (!len)

   184     return 0;

   185 

   186   for (i = 0; i < len; i++)

   187     if (!c->font->get_glyph (decomposed[i], 0, &glyphs[i]))

   188       return 0;

   189 

   190   for (i = 0; i < len; i++)

   191     output_char (c->buffer, decomposed[i], glyphs[i]);

   192 

   193   return len;

   194 }

   195 

   196 static inline void

   197 decompose_current_character (const hb_ot_shape_normalize_context_t *c, bool shortest)

   198 {

   199   hb_buffer_t * const buffer = c->buffer;

   200   hb_codepoint_t u = buffer->cur().codepoint;

   201   hb_codepoint_t glyph;

   202 

   203   /* Kind of a cute waterfall here... */

   204   if (shortest && c->font->get_glyph (u, 0, &glyph))

   205     next_char (buffer, glyph);

   206   else if (decompose (c, shortest, u))

   207     skip_char (buffer);

   208   else if (!shortest && c->font->get_glyph (u, 0, &glyph))

   209     next_char (buffer, glyph);

   210   else if (decompose_compatibility (c, u))

   211     skip_char (buffer);

   212   else

   213     next_char (buffer, glyph); /* glyph is initialized in earlier branches. */

   214 }

   215 

   216 static inline void

   217 handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)

   218 {

   219   /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */

   220   hb_buffer_t * const buffer = c->buffer;

   221   hb_font_t * const font = c->font;

   222   for (; buffer->idx < end - 1;) {

   223     if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {

   224       /* The next two lines are some ugly lines... But work. */

   225       if (font->get_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()))

   226       {

   227         buffer->replace_glyphs (2, 1, &buffer->cur().codepoint);

   228       }

   229       else

   230       {

   231         /* Just pass on the two characters separately, let GSUB do its magic. */

   232         set_glyph (buffer->cur(), font);

   233         buffer->next_glyph ();

   234         set_glyph (buffer->cur(), font);

   235         buffer->next_glyph ();

   236       }

   237       /* Skip any further variation selectors. */

   238       while (buffer->idx < end && unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint)))

   239       {

   240         set_glyph (buffer->cur(), font);

   241         buffer->next_glyph ();

   242       }

   243     } else {

   244       set_glyph (buffer->cur(), font);

   245       buffer->next_glyph ();

   246     }

   247   }

   248   if (likely (buffer->idx < end)) {

   249     set_glyph (buffer->cur(), font);

   250     buffer->next_glyph ();

   251   }

   252 }

   253 

   254 static inline void

   255 decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)

   256 {

   257   hb_buffer_t * const buffer = c->buffer;

   258   for (unsigned int i = buffer->idx; i < end; i++)

   259     if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {

   260       handle_variation_selector_cluster (c, end, short_circuit);

   261       return;

   262     }

   263 

   264   while (buffer->idx < end)

   265     decompose_current_character (c, short_circuit);

   266 }

   267 

   268 static inline void

   269 decompose_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool might_short_circuit, bool always_short_circuit)

   270 {

   271   if (likely (c->buffer->idx + 1 == end))

   272     decompose_current_character (c, might_short_circuit);

   273   else

   274     decompose_multi_char_cluster (c, end, always_short_circuit);

   275 }

   276 

   277 

   278 static int

   279 compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)

   280 {

   281   unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);

   282   unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);

   283 

   284   return a < b ? -1 : a == b ? 0 : +1;

   285 }

   286 

   287 

   288 void

   289 _hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,

   290                         hb_buffer_t *buffer,

   291                         hb_font_t *font)

   292 {

   293   if (unlikely (!buffer->len)) return;

   294 

   295   _hb_buffer_assert_unicode_vars (buffer);

   296 

   297   hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference;

   298   const hb_ot_shape_normalize_context_t c = {

   299     plan,

   300     buffer,

   301     font,

   302     buffer->unicode,

   303     plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode,

   304     plan->shaper->compose   ? plan->shaper->compose   : compose_unicode

   305   };

   306 

   307   bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE;

   308   bool might_short_circuit = always_short_circuit ||

   309                              (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED &&

   310                               mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT);

   311   unsigned int count;

   312 

   313   /* We do a fairly straightforward yet custom normalization process in three

   314    * separate rounds: decompose, reorder, recompose (if desired).  Currently

   315    * this makes two buffer swaps.  We can make it faster by moving the last

   316    * two rounds into the inner loop for the first round, but it's more readable

   317    * this way. */

   318 

   319 

   320   /* First round, decompose */

   321 

   322   buffer->clear_output ();

   323   count = buffer->len;

   324   for (buffer->idx = 0; buffer->idx < count;)

   325   {

   326     unsigned int end;

   327     for (end = buffer->idx + 1; end < count; end++)

   328       if (likely (!HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->info[end]))))

   329         break;

   330 

   331     decompose_cluster (&c, end, might_short_circuit, always_short_circuit);

   332   }

   333   buffer->swap_buffers ();

   334 

   335 

   336   /* Second round, reorder (inplace) */

   337 

   338   count = buffer->len;

   339   for (unsigned int i = 0; i < count; i++)

   340   {

   341     if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0)

   342       continue;

   343 

   344     unsigned int end;

   345     for (end = i + 1; end < count; end++)

   346       if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0)

   347         break;

   348 

   349     /* We are going to do a O(n^2).  Only do this if the sequence is short. */

   350     if (end - i > 10) {

   351       i = end;

   352       continue;

   353     }

   354 

   355     buffer->sort (i, end, compare_combining_class);

   356 

   357     i = end;

   358   }

   359 

   360 

   361   if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE ||

   362       mode == HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED)

   363     return;

   364 

   365   /* Third round, recompose */

   366 

   367   /* As noted in the comment earlier, we don't try to combine

   368    * ccc=0 chars with their previous Starter. */

   369 

   370   buffer->clear_output ();

   371   count = buffer->len;

   372   unsigned int starter = 0;

   373   buffer->next_glyph ();

   374   while (buffer->idx < count)

   375   {

   376     hb_codepoint_t composed, glyph;

   377     if (/* We don't try to compose a non-mark character with it's preceding starter.

   378          * This is both an optimization to avoid trying to compose every two neighboring

   379          * glyphs in most scripts AND a desired feature for Hangul.  Apparently Hangul

   380          * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */

   381         HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->cur())) &&

   382         /* If there's anything between the starter and this char, they should have CCC

   383          * smaller than this character's. */

   384         (starter == buffer->out_len - 1 ||

   385          _hb_glyph_info_get_modified_combining_class (&buffer->prev()) < _hb_glyph_info_get_modified_combining_class (&buffer->cur())) &&

   386         /* And compose. */

   387         c.compose (&c,

   388                    buffer->out_info[starter].codepoint,

   389                    buffer->cur().codepoint,

   390                    &composed) &&

   391         /* And the font has glyph for the composite. */

   392         font->get_glyph (composed, 0, &glyph))

   393     {

   394       /* Composes. */

   395       buffer->next_glyph (); /* Copy to out-buffer. */

   396       if (unlikely (buffer->in_error))

   397         return;

   398       buffer->merge_out_clusters (starter, buffer->out_len);

   399       buffer->out_len--; /* Remove the second composable. */

   400       /* Modify starter and carry on. */

   401       buffer->out_info[starter].codepoint = composed;

   402       buffer->out_info[starter].glyph_index() = glyph;

   403       _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer->unicode);

   404 

   405       continue;

   406     }

   407 

   408     /* Blocked, or doesn't compose. */

   409     buffer->next_glyph ();

   410 

   411     if (_hb_glyph_info_get_modified_combining_class (&buffer->prev()) == 0)

   412       starter = buffer->out_len - 1;

   413   }

   414   buffer->swap_buffers ();

   415 

   416 }