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 glyph; 201 202 /* Kind of a cute waterfall here... */ 203 if (shortest && c->font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 204 next_char (buffer, glyph); 205 else if (decompose (c, shortest, buffer->cur().codepoint)) 206 skip_char (buffer); 207 else if (!shortest && c->font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 208 next_char (buffer, glyph); 209 else if (decompose_compatibility (c, buffer->cur().codepoint)) 210 skip_char (buffer); 211 else 212 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */ 213 } 214 215 static inline void 216 handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit) 217 { 218 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */ 219 hb_buffer_t * const buffer = c->buffer; 220 hb_font_t * const font = c->font; 221 for (; buffer->idx < end - 1;) { 222 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) { 223 /* The next two lines are some ugly lines... But work. */ 224 if (font->get_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index())) 225 { 226 buffer->replace_glyphs (2, 1, &buffer->cur().codepoint); 227 } 228 else 229 { 230 /* Just pass on the two characters separately, let GSUB do its magic. */ 231 set_glyph (buffer->cur(), font); 232 buffer->next_glyph (); 233 set_glyph (buffer->cur(), font); 234 buffer->next_glyph (); 235 } 236 /* Skip any further variation selectors. */ 237 while (buffer->idx < end && unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint))) 238 { 239 set_glyph (buffer->cur(), font); 240 buffer->next_glyph (); 241 } 242 } else { 243 set_glyph (buffer->cur(), font); 244 buffer->next_glyph (); 245 } 246 } 247 if (likely (buffer->idx < end)) { 248 set_glyph (buffer->cur(), font); 249 buffer->next_glyph (); 250 } 251 } 252 253 static inline void 254 decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit) 255 { 256 hb_buffer_t * const buffer = c->buffer; 257 for (unsigned int i = buffer->idx; i < end; i++) 258 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) { 259 handle_variation_selector_cluster (c, end, short_circuit); 260 return; 261 } 262 263 while (buffer->idx < end) 264 decompose_current_character (c, short_circuit); 265 } 266 267 static inline void 268 decompose_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool might_short_circuit, bool always_short_circuit) 269 { 270 if (likely (c->buffer->idx + 1 == end)) 271 decompose_current_character (c, might_short_circuit); 272 else 273 decompose_multi_char_cluster (c, end, always_short_circuit); 274 } 275 276 277 static int 278 compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb) 279 { 280 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa); 281 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb); 282 283 return a < b ? -1 : a == b ? 0 : +1; 284 } 285 286 287 void 288 _hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan, 289 hb_buffer_t *buffer, 290 hb_font_t *font) 291 { 292 _hb_buffer_assert_unicode_vars (buffer); 293 294 hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference; 295 const hb_ot_shape_normalize_context_t c = { 296 plan, 297 buffer, 298 font, 299 buffer->unicode, 300 plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode, 301 plan->shaper->compose ? plan->shaper->compose : compose_unicode 302 }; 303 304 bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE; 305 bool might_short_circuit = always_short_circuit || 306 (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED && 307 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT); 308 unsigned int count; 309 310 /* We do a fairly straightforward yet custom normalization process in three 311 * separate rounds: decompose, reorder, recompose (if desired). Currently 312 * this makes two buffer swaps. We can make it faster by moving the last 313 * two rounds into the inner loop for the first round, but it's more readable 314 * this way. */ 315 316 317 /* First round, decompose */ 318 319 buffer->clear_output (); 320 count = buffer->len; 321 for (buffer->idx = 0; buffer->idx < count;) 322 { 323 unsigned int end; 324 for (end = buffer->idx + 1; end < count; end++) 325 if (buffer->cur().cluster != buffer->info[end].cluster) 326 break; 327 328 decompose_cluster (&c, end, might_short_circuit, always_short_circuit); 329 } 330 buffer->swap_buffers (); 331 332 333 /* Second round, reorder (inplace) */ 334 335 count = buffer->len; 336 for (unsigned int i = 0; i < count; i++) 337 { 338 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0) 339 continue; 340 341 unsigned int end; 342 for (end = i + 1; end < count; end++) 343 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0) 344 break; 345 346 /* We are going to do a bubble-sort. Only do this if the 347 * sequence is short. Doing it on long sequences can result 348 * in an O(n^2) DoS. */ 349 if (end - i > 10) { 350 i = end; 351 continue; 352 } 353 354 hb_bubble_sort (buffer->info + i, end - i, compare_combining_class); 355 356 i = end; 357 } 358 359 360 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE || 361 mode == HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED) 362 return; 363 364 /* Third round, recompose */ 365 366 /* As noted in the comment earlier, we don't try to combine 367 * ccc=0 chars with their previous Starter. */ 368 369 buffer->clear_output (); 370 count = buffer->len; 371 unsigned int starter = 0; 372 buffer->next_glyph (); 373 while (buffer->idx < count) 374 { 375 hb_codepoint_t composed, glyph; 376 if (/* We don't try to compose a non-mark character with it's preceding starter. 377 * This is both an optimization to avoid trying to compose every two neighboring 378 * glyphs in most scripts AND a desired feature for Hangul. Apparently Hangul 379 * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */ 380 HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->cur())) && 381 /* If there's anything between the starter and this char, they should have CCC 382 * smaller than this character's. */ 383 (starter == buffer->out_len - 1 || 384 _hb_glyph_info_get_modified_combining_class (&buffer->prev()) < _hb_glyph_info_get_modified_combining_class (&buffer->cur())) && 385 /* And compose. */ 386 c.compose (&c, 387 buffer->out_info[starter].codepoint, 388 buffer->cur().codepoint, 389 &composed) && 390 /* And the font has glyph for the composite. */ 391 font->get_glyph (composed, 0, &glyph)) 392 { 393 /* Composes. */ 394 buffer->next_glyph (); /* Copy to out-buffer. */ 395 if (unlikely (buffer->in_error)) 396 return; 397 buffer->merge_out_clusters (starter, buffer->out_len); 398 buffer->out_len--; /* Remove the second composable. */ 399 /* Modify starter and carry on. */ 400 buffer->out_info[starter].codepoint = composed; 401 buffer->out_info[starter].glyph_index() = glyph; 402 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer->unicode); 403 404 continue; 405 } 406 407 /* Blocked, or doesn't compose. */ 408 buffer->next_glyph (); 409 410 if (_hb_glyph_info_get_modified_combining_class (&buffer->prev()) == 0) 411 starter = buffer->out_len - 1; 412 } 413 buffer->swap_buffers (); 414 415 } 416