hb-ot-shape-normalize.cc revision 5821806d5e7f356e8fa4b058a389a808ea183019
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-private.hh" 29 30 31/* 32 * HIGHLEVEL DESIGN: 33 * 34 * This file exports one main function: _hb_ot_shape_normalize(). 35 * 36 * This function closely reflects the Unicode Normalization Algorithm, 37 * yet it's different. 38 * 39 * Each shaper specifies whether it prefers decomposed (NFD) or composed (NFC). 40 * The logic however tries to use whatever the font can support. 41 * 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. 46 * 47 * The goals are: 48 * 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. 55 * 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. 59 * 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. 63 * 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. 69 * 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. 73 * 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: 80 * 81 * egrep "`echo -n ';('; grep ';<' UnicodeData.txt | cut -d';' -f1 | tr '\n' '|'; echo ') '`" UnicodeData.txt 82 */ 83 84static hb_bool_t 85decompose_func (hb_unicode_funcs_t *unicode, 86 hb_codepoint_t ab, 87 hb_codepoint_t *a, 88 hb_codepoint_t *b) 89{ 90 /* XXX FIXME, move these to complex shapers and propagage to normalizer.*/ 91 switch (ab) { 92 case 0x0AC9 : return false; 93 94 case 0x0931 : return false; 95 case 0x0B94 : return false; 96 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; 103 104 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; 116#if 0 117 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; 121#endif 122 } 123 return unicode->decompose (ab, a, b); 124} 125 126static hb_bool_t 127compose_func (hb_unicode_funcs_t *unicode, 128 hb_codepoint_t a, 129 hb_codepoint_t b, 130 hb_codepoint_t *ab) 131{ 132 /* XXX, this belongs to indic normalizer. */ 133 if (HB_UNICODE_GENERAL_CATEGORY_IS_MARK (unicode->general_category (a))) 134 return false; 135 /* XXX, add composition-exclusion exceptions to Indic shaper. */ 136 if (a == 0x09AF && b == 0x09BC) { *ab = 0x09DF; return true; } 137 138 /* XXX, these belong to the hebew / default shaper. */ 139 /* Hebrew presentation-form shaping. 140 * https://bugzilla.mozilla.org/show_bug.cgi?id=728866 */ 141 // Hebrew presentation forms with dagesh, for characters 0x05D0..0x05EA; 142 // note that some letters do not have a dagesh presForm encoded 143 static const hb_codepoint_t sDageshForms[0x05EA - 0x05D0 + 1] = { 144 0xFB30, // ALEF 145 0xFB31, // BET 146 0xFB32, // GIMEL 147 0xFB33, // DALET 148 0xFB34, // HE 149 0xFB35, // VAV 150 0xFB36, // ZAYIN 151 0, // HET 152 0xFB38, // TET 153 0xFB39, // YOD 154 0xFB3A, // FINAL KAF 155 0xFB3B, // KAF 156 0xFB3C, // LAMED 157 0, // FINAL MEM 158 0xFB3E, // MEM 159 0, // FINAL NUN 160 0xFB40, // NUN 161 0xFB41, // SAMEKH 162 0, // AYIN 163 0xFB43, // FINAL PE 164 0xFB44, // PE 165 0, // FINAL TSADI 166 0xFB46, // TSADI 167 0xFB47, // QOF 168 0xFB48, // RESH 169 0xFB49, // SHIN 170 0xFB4A // TAV 171 }; 172 173 hb_bool_t found = unicode->compose (a, b, ab); 174 175 if (!found && (b & ~0x7F) == 0x0580) { 176 // special-case Hebrew presentation forms that are excluded from 177 // standard normalization, but wanted for old fonts 178 switch (b) { 179 case 0x05B4: // HIRIQ 180 if (a == 0x05D9) { // YOD 181 *ab = 0xFB1D; 182 found = true; 183 } 184 break; 185 case 0x05B7: // patah 186 if (a == 0x05F2) { // YIDDISH YOD YOD 187 *ab = 0xFB1F; 188 found = true; 189 } else if (a == 0x05D0) { // ALEF 190 *ab = 0xFB2E; 191 found = true; 192 } 193 break; 194 case 0x05B8: // QAMATS 195 if (a == 0x05D0) { // ALEF 196 *ab = 0xFB2F; 197 found = true; 198 } 199 break; 200 case 0x05B9: // HOLAM 201 if (a == 0x05D5) { // VAV 202 *ab = 0xFB4B; 203 found = true; 204 } 205 break; 206 case 0x05BC: // DAGESH 207 if (a >= 0x05D0 && a <= 0x05EA) { 208 *ab = sDageshForms[a - 0x05D0]; 209 found = (*ab != 0); 210 } else if (a == 0xFB2A) { // SHIN WITH SHIN DOT 211 *ab = 0xFB2C; 212 found = true; 213 } else if (a == 0xFB2B) { // SHIN WITH SIN DOT 214 *ab = 0xFB2D; 215 found = true; 216 } 217 break; 218 case 0x05BF: // RAFE 219 switch (a) { 220 case 0x05D1: // BET 221 *ab = 0xFB4C; 222 found = true; 223 break; 224 case 0x05DB: // KAF 225 *ab = 0xFB4D; 226 found = true; 227 break; 228 case 0x05E4: // PE 229 *ab = 0xFB4E; 230 found = true; 231 break; 232 } 233 break; 234 case 0x05C1: // SHIN DOT 235 if (a == 0x05E9) { // SHIN 236 *ab = 0xFB2A; 237 found = true; 238 } else if (a == 0xFB49) { // SHIN WITH DAGESH 239 *ab = 0xFB2C; 240 found = true; 241 } 242 break; 243 case 0x05C2: // SIN DOT 244 if (a == 0x05E9) { // SHIN 245 *ab = 0xFB2B; 246 found = true; 247 } else if (a == 0xFB49) { // SHIN WITH DAGESH 248 *ab = 0xFB2D; 249 found = true; 250 } 251 break; 252 } 253 } 254 255 return found; 256} 257 258 259static inline void 260set_glyph (hb_glyph_info_t &info, hb_font_t *font) 261{ 262 font->get_glyph (info.codepoint, 0, &info.glyph_index()); 263} 264 265static inline void 266output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph) 267{ 268 buffer->cur().glyph_index() = glyph; 269 buffer->output_glyph (unichar); 270 _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer->unicode); 271} 272 273static inline void 274next_char (hb_buffer_t *buffer, hb_codepoint_t glyph) 275{ 276 buffer->cur().glyph_index() = glyph; 277 buffer->next_glyph (); 278} 279 280static inline void 281skip_char (hb_buffer_t *buffer) 282{ 283 buffer->skip_glyph (); 284} 285 286/* Returns 0 if didn't decompose, number of resulting characters otherwise. */ 287static inline unsigned int 288decompose (hb_font_t *font, hb_buffer_t *buffer, bool shortest, hb_codepoint_t ab) 289{ 290 hb_codepoint_t a, b, a_glyph, b_glyph; 291 292 if (!decompose_func (buffer->unicode, ab, &a, &b) || 293 (b && !font->get_glyph (b, 0, &b_glyph))) 294 return 0; 295 296 bool has_a = font->get_glyph (a, 0, &a_glyph); 297 if (shortest && has_a) { 298 /* Output a and b */ 299 output_char (buffer, a, a_glyph); 300 if (likely (b)) { 301 output_char (buffer, b, b_glyph); 302 return 2; 303 } 304 return 1; 305 } 306 307 unsigned int ret; 308 if ((ret = decompose (font, buffer, shortest, a))) { 309 if (b) { 310 output_char (buffer, b, b_glyph); 311 return ret + 1; 312 } 313 return ret; 314 } 315 316 if (has_a) { 317 output_char (buffer, a, a_glyph); 318 if (likely (b)) { 319 output_char (buffer, b, b_glyph); 320 return 2; 321 } 322 return 1; 323 } 324 325 return 0; 326} 327 328/* Returns 0 if didn't decompose, number of resulting characters otherwise. */ 329static inline bool 330decompose_compatibility (hb_font_t *font, hb_buffer_t *buffer, hb_codepoint_t u) 331{ 332 unsigned int len, i; 333 hb_codepoint_t decomposed[HB_UNICODE_MAX_DECOMPOSITION_LEN]; 334 hb_codepoint_t glyphs[HB_UNICODE_MAX_DECOMPOSITION_LEN]; 335 336 len = buffer->unicode->decompose_compatibility (u, decomposed); 337 if (!len) 338 return 0; 339 340 for (i = 0; i < len; i++) 341 if (!font->get_glyph (decomposed[i], 0, &glyphs[i])) 342 return 0; 343 344 for (i = 0; i < len; i++) 345 output_char (buffer, decomposed[i], glyphs[i]); 346 347 return len; 348} 349 350/* Returns true if recomposition may be benefitial. */ 351static inline bool 352decompose_current_character (hb_font_t *font, hb_buffer_t *buffer, bool shortest) 353{ 354 hb_codepoint_t glyph; 355 unsigned int len = 1; 356 357 /* Kind of a cute waterfall here... */ 358 if (shortest && font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 359 next_char (buffer, glyph); 360 else if ((len = decompose (font, buffer, shortest, buffer->cur().codepoint))) 361 skip_char (buffer); 362 else if (!shortest && font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 363 next_char (buffer, glyph); 364 else if ((len = decompose_compatibility (font, buffer, buffer->cur().codepoint))) 365 skip_char (buffer); 366 else 367 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */ 368 369 /* 370 * A recomposition would only be useful if we decomposed into at least three 371 * characters... 372 */ 373 return len > 2; 374} 375 376static inline void 377handle_variation_selector_cluster (hb_font_t *font, hb_buffer_t *buffer, unsigned int end) 378{ 379 for (; buffer->idx < end - 1;) { 380 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) { 381 /* The next two lines are some ugly lines... But work. */ 382 font->get_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()); 383 buffer->replace_glyphs (2, 1, &buffer->cur().codepoint); 384 } else { 385 set_glyph (buffer->cur(), font); 386 buffer->next_glyph (); 387 } 388 } 389 if (likely (buffer->idx < end)) { 390 set_glyph (buffer->cur(), font); 391 buffer->next_glyph (); 392 } 393} 394 395/* Returns true if recomposition may be benefitial. */ 396static inline bool 397decompose_multi_char_cluster (hb_font_t *font, hb_buffer_t *buffer, unsigned int end) 398{ 399 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */ 400 for (unsigned int i = buffer->idx; i < end; i++) 401 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) { 402 handle_variation_selector_cluster (font, buffer, end); 403 return false; 404 } 405 406 while (buffer->idx < end) 407 decompose_current_character (font, buffer, false); 408 /* We can be smarter here and only return true if there are at least two ccc!=0 marks. 409 * But does not matter. */ 410 return true; 411} 412 413static inline bool 414decompose_cluster (hb_font_t *font, hb_buffer_t *buffer, bool short_circuit, unsigned int end) 415{ 416 if (likely (buffer->idx + 1 == end)) 417 return decompose_current_character (font, buffer, short_circuit); 418 else 419 return decompose_multi_char_cluster (font, buffer, end); 420} 421 422 423static int 424compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb) 425{ 426 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa); 427 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb); 428 429 return a < b ? -1 : a == b ? 0 : +1; 430} 431 432 433void 434_hb_ot_shape_normalize (hb_font_t *font, hb_buffer_t *buffer, 435 hb_ot_shape_normalization_mode_t mode) 436{ 437 bool short_circuit = mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED && 438 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT; 439 bool can_use_recompose = false; 440 unsigned int count; 441 442 /* We do a fairly straightforward yet custom normalization process in three 443 * separate rounds: decompose, reorder, recompose (if desired). Currently 444 * this makes two buffer swaps. We can make it faster by moving the last 445 * two rounds into the inner loop for the first round, but it's more readable 446 * this way. */ 447 448 449 /* First round, decompose */ 450 451 buffer->clear_output (); 452 count = buffer->len; 453 for (buffer->idx = 0; buffer->idx < count;) 454 { 455 unsigned int end; 456 for (end = buffer->idx + 1; end < count; end++) 457 if (buffer->cur().cluster != buffer->info[end].cluster) 458 break; 459 460 can_use_recompose = decompose_cluster (font, buffer, short_circuit, end) || can_use_recompose; 461 } 462 buffer->swap_buffers (); 463 464 465 if (mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_FULL && !can_use_recompose) 466 return; /* Done! */ 467 468 469 /* Second round, reorder (inplace) */ 470 471 count = buffer->len; 472 for (unsigned int i = 0; i < count; i++) 473 { 474 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0) 475 continue; 476 477 unsigned int end; 478 for (end = i + 1; end < count; end++) 479 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0) 480 break; 481 482 /* We are going to do a bubble-sort. Only do this if the 483 * sequence is short. Doing it on long sequences can result 484 * in an O(n^2) DoS. */ 485 if (end - i > 10) { 486 i = end; 487 continue; 488 } 489 490 hb_bubble_sort (buffer->info + i, end - i, compare_combining_class); 491 492 i = end; 493 } 494 495 496 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED) 497 return; 498 499 /* Third round, recompose */ 500 501 /* As noted in the comment earlier, we don't try to combine 502 * ccc=0 chars with their previous Starter. */ 503 504 buffer->clear_output (); 505 count = buffer->len; 506 unsigned int starter = 0; 507 buffer->next_glyph (); 508 while (buffer->idx < count) 509 { 510 hb_codepoint_t composed, glyph; 511 if (/* If mode is NOT COMPOSED_FULL (ie. it's COMPOSED_DIACRITICS), we don't try to 512 * compose a CCC=0 character with it's preceding starter. */ 513 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_FULL || 514 _hb_glyph_info_get_modified_combining_class (&buffer->cur()) != 0) && 515 /* If there's anything between the starter and this char, they should have CCC 516 * smaller than this character's. */ 517 (starter == buffer->out_len - 1 || 518 _hb_glyph_info_get_modified_combining_class (&buffer->prev()) < _hb_glyph_info_get_modified_combining_class (&buffer->cur())) && 519 /* And compose. */ 520 compose_func (buffer->unicode, 521 buffer->out_info[starter].codepoint, 522 buffer->cur().codepoint, 523 &composed) && 524 /* And the font has glyph for the composite. */ 525 font->get_glyph (composed, 0, &glyph)) 526 { 527 /* Composes. */ 528 buffer->next_glyph (); /* Copy to out-buffer. */ 529 if (unlikely (buffer->in_error)) 530 return; 531 buffer->merge_out_clusters (starter, buffer->out_len); 532 buffer->out_len--; /* Remove the second composable. */ 533 buffer->out_info[starter].codepoint = composed; /* Modify starter and carry on. */ 534 set_glyph (buffer->out_info[starter], font); 535 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer->unicode); 536 537 continue; 538 } 539 540 /* Blocked, or doesn't compose. */ 541 buffer->next_glyph (); 542 543 if (_hb_glyph_info_get_modified_combining_class (&buffer->prev()) == 0) 544 starter = buffer->out_len - 1; 545 } 546 buffer->swap_buffers (); 547 548} 549