2 * Copyright © 2011,2012 Google, Inc.
4 * This is part of HarfBuzz, a text shaping library.
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.
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
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.
24 * Google Author(s): Behdad Esfahbod
27 #include "hb-ot-shape-normalize-private.hh"
28 #include "hb-ot-shape-private.hh"
34 * This file exports one main function: _hb_ot_shape_normalize().
36 * This function closely reflects the Unicode Normalization Algorithm,
39 * Each shaper specifies whether it prefers decomposed (NFD) or composed (NFC).
40 * The logic however tries to use whatever the font can support.
42 * In general what happens is that: each grapheme is decomposed in a chain
43 * of 1:2 decompositions, marks reordered, and then recomposed if desired,
44 * so far it's like Unicode Normalization. However, the decomposition and
45 * recomposition only happens if the font supports the resulting characters.
49 * - Try to render all canonically equivalent strings similarly. To really
50 * achieve this we have to always do the full decomposition and then
51 * selectively recompose from there. It's kinda too expensive though, so
52 * we skip some cases. For example, if composed is desired, we simply
53 * don't touch 1-character clusters that are supported by the font, even
54 * though their NFC may be different.
56 * - When a font has a precomposed character for a sequence but the 'ccmp'
57 * feature in the font is not adequate, use the precomposed character
58 * which typically has better mark positioning.
60 * - When a font does not support a combining mark, but supports it precomposed
61 * with previous base, use that. This needs the itemizer to have this
62 * knowledge too. We need to provide assistance to the itemizer.
64 * - When a font does not support a character but supports its decomposition,
65 * well, use the decomposition (preferring the canonical decomposition, but
66 * falling back to the compatibility decomposition if necessary). The
67 * compatibility decomposition is really nice to have, for characters like
68 * ellipsis, or various-sized space characters.
70 * - The complex shapers can customize the compose and decompose functions to
71 * offload some of their requirements to the normalizer. For example, the
72 * Indic shaper may want to disallow recomposing of two matras.
74 * - We try compatibility decomposition if decomposing through canonical
75 * decomposition alone failed to find a sequence that the font supports.
76 * We don't try compatibility decomposition recursively during the canonical
77 * decomposition phase. This has minimal impact. There are only a handful
78 * of Greek letter that have canonical decompositions that include characters
79 * with compatibility decomposition. Those can be found using this command:
81 * egrep "`echo -n ';('; grep ';<' UnicodeData.txt | cut -d';' -f1 | tr '\n' '|'; echo ') '`" UnicodeData.txt
85 decompose_func (hb_unicode_funcs_t *unicode,
90 /* XXX FIXME, move these to complex shapers and propagage to normalizer.*/
92 case 0x0AC9 : return false;
94 case 0x0931 : return false;
95 case 0x0B94 : return false;
97 /* These ones have Unicode decompositions, but we do it
98 * this way to be close to what Uniscribe does. */
99 case 0x0DDA : *a = 0x0DD9; *b= 0x0DDA; return true;
100 case 0x0DDC : *a = 0x0DD9; *b= 0x0DDC; return true;
101 case 0x0DDD : *a = 0x0DD9; *b= 0x0DDD; return true;
102 case 0x0DDE : *a = 0x0DD9; *b= 0x0DDE; return true;
104 case 0x0F77 : *a = 0x0FB2; *b= 0x0F81; return true;
105 case 0x0F79 : *a = 0x0FB3; *b= 0x0F81; return true;
106 case 0x17BE : *a = 0x17C1; *b= 0x17BE; return true;
107 case 0x17BF : *a = 0x17C1; *b= 0x17BF; return true;
108 case 0x17C0 : *a = 0x17C1; *b= 0x17C0; return true;
109 case 0x17C4 : *a = 0x17C1; *b= 0x17C4; return true;
110 case 0x17C5 : *a = 0x17C1; *b= 0x17C5; return true;
111 case 0x1925 : *a = 0x1920; *b= 0x1923; return true;
112 case 0x1926 : *a = 0x1920; *b= 0x1924; return true;
113 case 0x1B3C : *a = 0x1B42; *b= 0x1B3C; return true;
114 case 0x1112E : *a = 0x11127; *b= 0x11131; return true;
115 case 0x1112F : *a = 0x11127; *b= 0x11132; return true;
117 case 0x0B57 : *a = 0xno decomp, -> RIGHT; return true;
118 case 0x1C29 : *a = 0xno decomp, -> LEFT; return true;
119 case 0xA9C0 : *a = 0xno decomp, -> RIGHT; return true;
120 case 0x111BF : *a = 0xno decomp, -> ABOVE; return true;
123 return unicode->decompose (ab, a, b);
127 compose_func (hb_unicode_funcs_t *unicode,
132 /* XXX, this belongs to indic normalizer. */
133 if ((FLAG (unicode->general_category (a)) &
134 (FLAG (HB_UNICODE_GENERAL_CATEGORY_SPACING_MARK) |
135 FLAG (HB_UNICODE_GENERAL_CATEGORY_ENCLOSING_MARK) |
136 FLAG (HB_UNICODE_GENERAL_CATEGORY_NON_SPACING_MARK))))
138 /* XXX, add composition-exclusion exceptions to Indic shaper. */
139 if (a == 0x09AF && b == 0x09BC) { *ab = 0x09DF; return true; }
141 /* XXX, these belong to the hebew / default shaper. */
142 /* Hebrew presentation-form shaping.
143 * https://bugzilla.mozilla.org/show_bug.cgi?id=728866 */
144 // Hebrew presentation forms with dagesh, for characters 0x05D0..0x05EA;
145 // note that some letters do not have a dagesh presForm encoded
146 static const hb_codepoint_t sDageshForms[0x05EA - 0x05D0 + 1] = {
176 hb_bool_t found = unicode->compose (a, b, ab);
178 if (!found && (b & ~0x7F) == 0x0580) {
179 // special-case Hebrew presentation forms that are excluded from
180 // standard normalization, but wanted for old fonts
182 case 0x05B4: // HIRIQ
183 if (a == 0x05D9) { // YOD
188 case 0x05B7: // patah
189 if (a == 0x05F2) { // YIDDISH YOD YOD
192 } else if (a == 0x05D0) { // ALEF
197 case 0x05B8: // QAMATS
198 if (a == 0x05D0) { // ALEF
203 case 0x05B9: // HOLAM
204 if (a == 0x05D5) { // VAV
209 case 0x05BC: // DAGESH
210 if (a >= 0x05D0 && a <= 0x05EA) {
211 *ab = sDageshForms[a - 0x05D0];
213 } else if (a == 0xFB2A) { // SHIN WITH SHIN DOT
216 } else if (a == 0xFB2B) { // SHIN WITH SIN DOT
237 case 0x05C1: // SHIN DOT
238 if (a == 0x05E9) { // SHIN
241 } else if (a == 0xFB49) { // SHIN WITH DAGESH
246 case 0x05C2: // SIN DOT
247 if (a == 0x05E9) { // SHIN
250 } else if (a == 0xFB49) { // SHIN WITH DAGESH
263 set_glyph (hb_glyph_info_t &info, hb_font_t *font)
265 font->get_glyph (info.codepoint, 0, &info.glyph_index());
269 output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph)
271 buffer->cur().glyph_index() = glyph;
272 buffer->output_glyph (unichar);
273 _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer->unicode);
277 next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)
279 buffer->cur().glyph_index() = glyph;
280 buffer->next_glyph ();
284 skip_char (hb_buffer_t *buffer)
286 buffer->skip_glyph ();
289 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */
290 static inline unsigned int
291 decompose (hb_font_t *font, hb_buffer_t *buffer, bool shortest, hb_codepoint_t ab)
293 hb_codepoint_t a, b, a_glyph, b_glyph;
295 if (!decompose_func (buffer->unicode, ab, &a, &b) ||
296 (b && !font->get_glyph (b, 0, &b_glyph)))
299 bool has_a = font->get_glyph (a, 0, &a_glyph);
300 if (shortest && has_a) {
302 output_char (buffer, a, a_glyph);
304 output_char (buffer, b, b_glyph);
311 if ((ret = decompose (font, buffer, shortest, a))) {
313 output_char (buffer, b, b_glyph);
320 output_char (buffer, a, a_glyph);
322 output_char (buffer, b, b_glyph);
331 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */
333 decompose_compatibility (hb_font_t *font, hb_buffer_t *buffer, hb_codepoint_t u)
336 hb_codepoint_t decomposed[HB_UNICODE_MAX_DECOMPOSITION_LEN];
337 hb_codepoint_t glyphs[HB_UNICODE_MAX_DECOMPOSITION_LEN];
339 len = buffer->unicode->decompose_compatibility (u, decomposed);
343 for (i = 0; i < len; i++)
344 if (!font->get_glyph (decomposed[i], 0, &glyphs[i]))
347 for (i = 0; i < len; i++)
348 output_char (buffer, decomposed[i], glyphs[i]);
353 /* Returns true if recomposition may be benefitial. */
355 decompose_current_character (hb_font_t *font, hb_buffer_t *buffer, bool shortest)
357 hb_codepoint_t glyph;
358 unsigned int len = 1;
360 /* Kind of a cute waterfall here... */
361 if (shortest && font->get_glyph (buffer->cur().codepoint, 0, &glyph))
362 next_char (buffer, glyph);
363 else if ((len = decompose (font, buffer, shortest, buffer->cur().codepoint)))
365 else if (!shortest && font->get_glyph (buffer->cur().codepoint, 0, &glyph))
366 next_char (buffer, glyph);
367 else if ((len = decompose_compatibility (font, buffer, buffer->cur().codepoint)))
370 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */
373 * A recomposition would only be useful if we decomposed into at least three
380 handle_variation_selector_cluster (hb_font_t *font, hb_buffer_t *buffer, unsigned int end)
382 for (; buffer->idx < end - 1;) {
383 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {
384 /* The next two lines are some ugly lines... But work. */
385 font->get_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index());
386 buffer->replace_glyphs (2, 1, &buffer->cur().codepoint);
388 set_glyph (buffer->cur(), font);
389 buffer->next_glyph ();
392 if (likely (buffer->idx < end)) {
393 set_glyph (buffer->cur(), font);
394 buffer->next_glyph ();
398 /* Returns true if recomposition may be benefitial. */
400 decompose_multi_char_cluster (hb_font_t *font, hb_buffer_t *buffer, unsigned int end)
402 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */
403 for (unsigned int i = buffer->idx; i < end; i++)
404 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {
405 handle_variation_selector_cluster (font, buffer, end);
409 while (buffer->idx < end)
410 decompose_current_character (font, buffer, false);
411 /* We can be smarter here and only return true if there are at least two ccc!=0 marks.
412 * But does not matter. */
417 decompose_cluster (hb_font_t *font, hb_buffer_t *buffer, bool recompose, unsigned int end)
419 if (likely (buffer->idx + 1 == end))
420 return decompose_current_character (font, buffer, recompose);
422 return decompose_multi_char_cluster (font, buffer, end);
427 compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
429 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);
430 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);
432 return a < b ? -1 : a == b ? 0 : +1;
437 _hb_ot_shape_normalize (hb_font_t *font, hb_buffer_t *buffer,
438 hb_ot_shape_normalization_mode_t mode)
440 bool recompose = mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED;
441 bool can_use_recompose = false;
444 /* We do a fairly straightforward yet custom normalization process in three
445 * separate rounds: decompose, reorder, recompose (if desired). Currently
446 * this makes two buffer swaps. We can make it faster by moving the last
447 * two rounds into the inner loop for the first round, but it's more readable
451 /* First round, decompose */
453 buffer->clear_output ();
455 for (buffer->idx = 0; buffer->idx < count;)
458 for (end = buffer->idx + 1; end < count; end++)
459 if (buffer->cur().cluster != buffer->info[end].cluster)
462 can_use_recompose = decompose_cluster (font, buffer, recompose, end) || can_use_recompose;
464 buffer->swap_buffers ();
467 if (mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_FULL && !can_use_recompose)
471 /* Second round, reorder (inplace) */
474 for (unsigned int i = 0; i < count; i++)
476 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0)
480 for (end = i + 1; end < count; end++)
481 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0)
484 /* We are going to do a bubble-sort. Only do this if the
485 * sequence is short. Doing it on long sequences can result
486 * in an O(n^2) DoS. */
492 hb_bubble_sort (buffer->info + i, end - i, compare_combining_class);
501 /* Third round, recompose */
503 /* As noted in the comment earlier, we don't try to combine
504 * ccc=0 chars with their previous Starter. */
506 buffer->clear_output ();
508 unsigned int starter = 0;
509 buffer->next_glyph ();
510 while (buffer->idx < count)
512 hb_codepoint_t composed, glyph;
513 if (/* If mode is NOT COMPOSED_FULL (ie. it's COMPOSED_DIACRITICS), we don't try to
514 * compose a CCC=0 character with it's preceding starter. */
515 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_FULL ||
516 _hb_glyph_info_get_modified_combining_class (&buffer->cur()) != 0) &&
517 /* If there's anything between the starter and this char, they should have CCC
518 * smaller than this character's. */
519 (starter == buffer->out_len - 1 ||
520 _hb_glyph_info_get_modified_combining_class (&buffer->prev()) < _hb_glyph_info_get_modified_combining_class (&buffer->cur())) &&
522 compose_func (buffer->unicode,
523 buffer->out_info[starter].codepoint,
524 buffer->cur().codepoint,
526 /* And the font has glyph for the composite. */
527 font->get_glyph (composed, 0, &glyph))
530 buffer->next_glyph (); /* Copy to out-buffer. */
531 if (unlikely (buffer->in_error))
533 buffer->merge_out_clusters (starter, buffer->out_len);
534 buffer->out_len--; /* Remove the second composable. */
535 buffer->out_info[starter].codepoint = composed; /* Modify starter and carry on. */
536 set_glyph (buffer->out_info[starter], font);
537 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer->unicode);
542 /* Blocked, or doesn't compose. */
543 buffer->next_glyph ();
545 if (_hb_glyph_info_get_modified_combining_class (&buffer->prev()) == 0)
546 starter = buffer->out_len - 1;
548 buffer->swap_buffers ();