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 85static bool 86decompose_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 94static bool 95compose_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 103static inline void 104set_glyph (hb_glyph_info_t &info, hb_font_t *font) 105{ 106 font->get_glyph (info.codepoint, 0, &info.glyph_index()); 107} 108 109static inline void 110output_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 117static inline void 118next_char (hb_buffer_t *buffer, hb_codepoint_t glyph) 119{ 120 buffer->cur().glyph_index() = glyph; 121 buffer->next_glyph (); 122} 123 124static inline void 125skip_char (hb_buffer_t *buffer) 126{ 127 buffer->skip_glyph (); 128} 129 130/* Returns 0 if didn't decompose, number of resulting characters otherwise. */ 131static inline unsigned int 132decompose (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; 135 136 if (!c->decompose (c, ab, &a, &b) || 137 (b && !c->font->get_glyph (b, 0, &b_glyph))) 138 return 0; 139 140 bool has_a = c->font->get_glyph (a, 0, &a_glyph); 141 if (shortest && has_a) { 142 /* Output a and b */ 143 output_char (c->buffer, a, a_glyph); 144 if (likely (b)) { 145 output_char (c->buffer, b, b_glyph); 146 return 2; 147 } 148 return 1; 149 } 150 151 unsigned int ret; 152 if ((ret = decompose (c, shortest, a))) { 153 if (b) { 154 output_char (c->buffer, b, b_glyph); 155 return ret + 1; 156 } 157 return ret; 158 } 159 160 if (has_a) { 161 output_char (c->buffer, a, a_glyph); 162 if (likely (b)) { 163 output_char (c->buffer, b, b_glyph); 164 return 2; 165 } 166 return 1; 167 } 168 169 return 0; 170} 171 172/* Returns 0 if didn't decompose, number of resulting characters otherwise. */ 173static inline bool 174decompose_compatibility (const hb_ot_shape_normalize_context_t *c, hb_codepoint_t u) 175{ 176 unsigned int len, i; 177 hb_codepoint_t decomposed[HB_UNICODE_MAX_DECOMPOSITION_LEN]; 178 hb_codepoint_t glyphs[HB_UNICODE_MAX_DECOMPOSITION_LEN]; 179 180 len = c->buffer->unicode->decompose_compatibility (u, decomposed); 181 if (!len) 182 return 0; 183 184 for (i = 0; i < len; i++) 185 if (!c->font->get_glyph (decomposed[i], 0, &glyphs[i])) 186 return 0; 187 188 for (i = 0; i < len; i++) 189 output_char (c->buffer, decomposed[i], glyphs[i]); 190 191 return len; 192} 193 194/* Returns true if recomposition may be benefitial. */ 195static inline void 196decompose_current_character (const hb_ot_shape_normalize_context_t *c, bool shortest) 197{ 198 hb_buffer_t * const buffer = c->buffer; 199 hb_codepoint_t glyph; 200 201 /* Kind of a cute waterfall here... */ 202 if (shortest && c->font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 203 next_char (buffer, glyph); 204 else if (decompose (c, shortest, buffer->cur().codepoint)) 205 skip_char (buffer); 206 else if (!shortest && c->font->get_glyph (buffer->cur().codepoint, 0, &glyph)) 207 next_char (buffer, glyph); 208 else if (decompose_compatibility (c, buffer->cur().codepoint)) 209 skip_char (buffer); 210 else 211 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */ 212} 213 214static inline void 215handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end) 216{ 217 hb_buffer_t * const buffer = c->buffer; 218 for (; buffer->idx < end - 1;) { 219 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) { 220 /* The next two lines are some ugly lines... But work. */ 221 c->font->get_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()); 222 buffer->replace_glyphs (2, 1, &buffer->cur().codepoint); 223 } else { 224 set_glyph (buffer->cur(), c->font); 225 buffer->next_glyph (); 226 } 227 } 228 if (likely (buffer->idx < end)) { 229 set_glyph (buffer->cur(), c->font); 230 buffer->next_glyph (); 231 } 232} 233 234/* Returns true if recomposition may be benefitial. */ 235static inline void 236decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end) 237{ 238 hb_buffer_t * const buffer = c->buffer; 239 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */ 240 for (unsigned int i = buffer->idx; i < end; i++) 241 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) { 242 handle_variation_selector_cluster (c, end); 243 return; 244 } 245 246 while (buffer->idx < end) 247 decompose_current_character (c, false); 248} 249 250static inline void 251decompose_cluster (const hb_ot_shape_normalize_context_t *c, bool short_circuit, unsigned int end) 252{ 253 if (likely (c->buffer->idx + 1 == end)) 254 decompose_current_character (c, short_circuit); 255 else 256 decompose_multi_char_cluster (c, end); 257} 258 259 260static int 261compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb) 262{ 263 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa); 264 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb); 265 266 return a < b ? -1 : a == b ? 0 : +1; 267} 268 269 270void 271_hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan, 272 hb_buffer_t *buffer, 273 hb_font_t *font) 274{ 275 hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference ? 276 plan->shaper->normalization_preference (&buffer->props) : 277 HB_OT_SHAPE_NORMALIZATION_MODE_DEFAULT; 278 const hb_ot_shape_normalize_context_t c = { 279 plan, 280 buffer, 281 font, 282 buffer->unicode, 283 plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode, 284 plan->shaper->compose ? plan->shaper->compose : compose_unicode 285 }; 286 287 bool short_circuit = mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED && 288 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT; 289 unsigned int count; 290 291 /* We do a fairly straightforward yet custom normalization process in three 292 * separate rounds: decompose, reorder, recompose (if desired). Currently 293 * this makes two buffer swaps. We can make it faster by moving the last 294 * two rounds into the inner loop for the first round, but it's more readable 295 * this way. */ 296 297 298 /* First round, decompose */ 299 300 buffer->clear_output (); 301 count = buffer->len; 302 for (buffer->idx = 0; buffer->idx < count;) 303 { 304 unsigned int end; 305 for (end = buffer->idx + 1; end < count; end++) 306 if (buffer->cur().cluster != buffer->info[end].cluster) 307 break; 308 309 decompose_cluster (&c, short_circuit, end); 310 } 311 buffer->swap_buffers (); 312 313 314 /* Second round, reorder (inplace) */ 315 316 count = buffer->len; 317 for (unsigned int i = 0; i < count; i++) 318 { 319 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0) 320 continue; 321 322 unsigned int end; 323 for (end = i + 1; end < count; end++) 324 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0) 325 break; 326 327 /* We are going to do a bubble-sort. Only do this if the 328 * sequence is short. Doing it on long sequences can result 329 * in an O(n^2) DoS. */ 330 if (end - i > 10) { 331 i = end; 332 continue; 333 } 334 335 hb_bubble_sort (buffer->info + i, end - i, compare_combining_class); 336 337 i = end; 338 } 339 340 341 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED) 342 return; 343 344 /* Third round, recompose */ 345 346 /* As noted in the comment earlier, we don't try to combine 347 * ccc=0 chars with their previous Starter. */ 348 349 buffer->clear_output (); 350 count = buffer->len; 351 unsigned int starter = 0; 352 buffer->next_glyph (); 353 while (buffer->idx < count) 354 { 355 hb_codepoint_t composed, glyph; 356 if (/* If mode is NOT COMPOSED_FULL (ie. it's COMPOSED_DIACRITICS), we don't try to 357 * compose a non-mark character with it's preceding starter. This is just an 358 * optimization to avoid trying to compose every two neighboring glyphs in most 359 * scripts. */ 360 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_FULL || 361 HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->cur()))) && 362 /* If there's anything between the starter and this char, they should have CCC 363 * smaller than this character's. */ 364 (starter == buffer->out_len - 1 || 365 _hb_glyph_info_get_modified_combining_class (&buffer->prev()) < _hb_glyph_info_get_modified_combining_class (&buffer->cur())) && 366 /* And compose. */ 367 c.compose (&c, 368 buffer->out_info[starter].codepoint, 369 buffer->cur().codepoint, 370 &composed) && 371 /* And the font has glyph for the composite. */ 372 font->get_glyph (composed, 0, &glyph)) 373 { 374 /* Composes. */ 375 buffer->next_glyph (); /* Copy to out-buffer. */ 376 if (unlikely (buffer->in_error)) 377 return; 378 buffer->merge_out_clusters (starter, buffer->out_len); 379 buffer->out_len--; /* Remove the second composable. */ 380 buffer->out_info[starter].codepoint = composed; /* Modify starter and carry on. */ 381 set_glyph (buffer->out_info[starter], font); 382 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer->unicode); 383 384 continue; 385 } 386 387 /* Blocked, or doesn't compose. */ 388 buffer->next_glyph (); 389 390 if (_hb_glyph_info_get_modified_combining_class (&buffer->prev()) == 0) 391 starter = buffer->out_len - 1; 392 } 393 buffer->swap_buffers (); 394 395} 396