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.hh"
28
29 #ifndef HB_NO_OT_SHAPE
30
31 #include "hb-ot-shape-normalize.hh"
32 #include "hb-ot-shape-complex.hh"
33 #include "hb-ot-shape.hh"
34
35
36 /*
37 * HIGHLEVEL DESIGN:
38 *
39 * This file exports one main function: _hb_ot_shape_normalize().
40 *
41 * This function closely reflects the Unicode Normalization Algorithm,
42 * yet it's different.
43 *
44 * Each shaper specifies whether it prefers decomposed (NFD) or composed (NFC).
45 * The logic however tries to use whatever the font can support.
46 *
47 * In general what happens is that: each grapheme is decomposed in a chain
48 * of 1:2 decompositions, marks reordered, and then recomposed if desired,
49 * so far it's like Unicode Normalization. However, the decomposition and
50 * recomposition only happens if the font supports the resulting characters.
51 *
52 * The goals are:
53 *
54 * - Try to render all canonically equivalent strings similarly. To really
55 * achieve this we have to always do the full decomposition and then
56 * selectively recompose from there. It's kinda too expensive though, so
57 * we skip some cases. For example, if composed is desired, we simply
58 * don't touch 1-character clusters that are supported by the font, even
59 * though their NFC may be different.
60 *
61 * - When a font has a precomposed character for a sequence but the 'ccmp'
62 * feature in the font is not adequate, use the precomposed character
63 * which typically has better mark positioning.
64 *
65 * - When a font does not support a combining mark, but supports it precomposed
66 * with previous base, use that. This needs the itemizer to have this
67 * knowledge too. We need to provide assistance to the itemizer.
68 *
69 * - When a font does not support a character but supports its canonical
70 * decomposition, well, use the decomposition.
71 *
72 * - The complex shapers can customize the compose and decompose functions to
73 * offload some of their requirements to the normalizer. For example, the
74 * Indic shaper may want to disallow recomposing of two matras.
75 */
76
77 static bool
decompose_unicode(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t ab,hb_codepoint_t * a,hb_codepoint_t * b)78 decompose_unicode (const hb_ot_shape_normalize_context_t *c,
79 hb_codepoint_t ab,
80 hb_codepoint_t *a,
81 hb_codepoint_t *b)
82 {
83 return (bool) c->unicode->decompose (ab, a, b);
84 }
85
86 static bool
compose_unicode(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t a,hb_codepoint_t b,hb_codepoint_t * ab)87 compose_unicode (const hb_ot_shape_normalize_context_t *c,
88 hb_codepoint_t a,
89 hb_codepoint_t b,
90 hb_codepoint_t *ab)
91 {
92 return (bool) c->unicode->compose (a, b, ab);
93 }
94
95 static inline void
set_glyph(hb_glyph_info_t & info,hb_font_t * font)96 set_glyph (hb_glyph_info_t &info, hb_font_t *font)
97 {
98 (void) font->get_nominal_glyph (info.codepoint, &info.glyph_index());
99 }
100
101 static inline void
output_char(hb_buffer_t * buffer,hb_codepoint_t unichar,hb_codepoint_t glyph)102 output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph)
103 {
104 /* This is very confusing indeed. */
105 buffer->cur().glyph_index() = glyph;
106 (void) buffer->output_glyph (unichar);
107 _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer);
108 }
109
110 static inline void
next_char(hb_buffer_t * buffer,hb_codepoint_t glyph)111 next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)
112 {
113 buffer->cur().glyph_index() = glyph;
114 (void) buffer->next_glyph ();
115 }
116
117 static inline void
skip_char(hb_buffer_t * buffer)118 skip_char (hb_buffer_t *buffer)
119 {
120 buffer->skip_glyph ();
121 }
122
123 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */
124 static inline unsigned int
decompose(const hb_ot_shape_normalize_context_t * c,bool shortest,hb_codepoint_t ab)125 decompose (const hb_ot_shape_normalize_context_t *c, bool shortest, hb_codepoint_t ab)
126 {
127 hb_codepoint_t a = 0, b = 0, a_glyph = 0, b_glyph = 0;
128 hb_buffer_t * const buffer = c->buffer;
129 hb_font_t * const font = c->font;
130
131 if (!c->decompose (c, ab, &a, &b) ||
132 (b && !font->get_nominal_glyph (b, &b_glyph)))
133 return 0;
134
135 bool has_a = (bool) font->get_nominal_glyph (a, &a_glyph);
136 if (shortest && has_a) {
137 /* Output a and b */
138 output_char (buffer, a, a_glyph);
139 if (likely (b)) {
140 output_char (buffer, b, b_glyph);
141 return 2;
142 }
143 return 1;
144 }
145
146 unsigned int ret;
147 if ((ret = decompose (c, shortest, a))) {
148 if (b) {
149 output_char (buffer, b, b_glyph);
150 return ret + 1;
151 }
152 return ret;
153 }
154
155 if (has_a) {
156 output_char (buffer, a, a_glyph);
157 if (likely (b)) {
158 output_char (buffer, b, b_glyph);
159 return 2;
160 }
161 return 1;
162 }
163
164 return 0;
165 }
166
167 static inline void
decompose_current_character(const hb_ot_shape_normalize_context_t * c,bool shortest)168 decompose_current_character (const hb_ot_shape_normalize_context_t *c, bool shortest)
169 {
170 hb_buffer_t * const buffer = c->buffer;
171 hb_codepoint_t u = buffer->cur().codepoint;
172 hb_codepoint_t glyph = 0;
173
174 if (shortest && c->font->get_nominal_glyph (u, &glyph, c->not_found))
175 {
176 next_char (buffer, glyph);
177 return;
178 }
179
180 if (decompose (c, shortest, u))
181 {
182 skip_char (buffer);
183 return;
184 }
185
186 if (!shortest && c->font->get_nominal_glyph (u, &glyph, c->not_found))
187 {
188 next_char (buffer, glyph);
189 return;
190 }
191
192 if (_hb_glyph_info_is_unicode_space (&buffer->cur()))
193 {
194 hb_codepoint_t space_glyph;
195 hb_unicode_funcs_t::space_t space_type = buffer->unicode->space_fallback_type (u);
196 if (space_type != hb_unicode_funcs_t::NOT_SPACE &&
197 (c->font->get_nominal_glyph (0x0020, &space_glyph) || (space_glyph = buffer->invisible)))
198 {
199 _hb_glyph_info_set_unicode_space_fallback_type (&buffer->cur(), space_type);
200 next_char (buffer, space_glyph);
201 buffer->scratch_flags |= HB_BUFFER_SCRATCH_FLAG_HAS_SPACE_FALLBACK;
202 return;
203 }
204 }
205
206 if (u == 0x2011u)
207 {
208 /* U+2011 is the only sensible character that is a no-break version of another character
209 * and not a space. The space ones are handled already. Handle this lone one. */
210 hb_codepoint_t other_glyph;
211 if (c->font->get_nominal_glyph (0x2010u, &other_glyph))
212 {
213 next_char (buffer, other_glyph);
214 return;
215 }
216 }
217
218 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */
219 }
220
221 static inline void
handle_variation_selector_cluster(const hb_ot_shape_normalize_context_t * c,unsigned int end,bool short_circuit HB_UNUSED)222 handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c,
223 unsigned int end,
224 bool short_circuit HB_UNUSED)
225 {
226 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */
227 hb_buffer_t * const buffer = c->buffer;
228 hb_font_t * const font = c->font;
229 for (; buffer->idx < end - 1 && buffer->successful;) {
230 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {
231 if (font->get_variation_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()))
232 {
233 hb_codepoint_t unicode = buffer->cur().codepoint;
234 (void) buffer->replace_glyphs (2, 1, &unicode);
235 }
236 else
237 {
238 /* Just pass on the two characters separately, let GSUB do its magic. */
239 set_glyph (buffer->cur(), font);
240 (void) buffer->next_glyph ();
241 set_glyph (buffer->cur(), font);
242 (void) buffer->next_glyph ();
243 }
244 /* Skip any further variation selectors. */
245 while (buffer->idx < end &&
246 buffer->successful &&
247 unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint)))
248 {
249 set_glyph (buffer->cur(), font);
250 (void) buffer->next_glyph ();
251 }
252 }
253 else
254 {
255 set_glyph (buffer->cur(), font);
256 (void) buffer->next_glyph ();
257 }
258 }
259 if (likely (buffer->idx < end))
260 {
261 set_glyph (buffer->cur(), font);
262 (void) buffer->next_glyph ();
263 }
264 }
265
266 static inline void
decompose_multi_char_cluster(const hb_ot_shape_normalize_context_t * c,unsigned int end,bool short_circuit)267 decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)
268 {
269 hb_buffer_t * const buffer = c->buffer;
270 for (unsigned int i = buffer->idx; i < end && buffer->successful; i++)
271 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {
272 handle_variation_selector_cluster (c, end, short_circuit);
273 return;
274 }
275
276 while (buffer->idx < end && buffer->successful)
277 decompose_current_character (c, short_circuit);
278 }
279
280
281 static int
compare_combining_class(const hb_glyph_info_t * pa,const hb_glyph_info_t * pb)282 compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
283 {
284 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);
285 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);
286
287 return a < b ? -1 : a == b ? 0 : +1;
288 }
289
290
291 void
_hb_ot_shape_normalize(const hb_ot_shape_plan_t * plan,hb_buffer_t * buffer,hb_font_t * font)292 _hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,
293 hb_buffer_t *buffer,
294 hb_font_t *font)
295 {
296 if (unlikely (!buffer->len)) return;
297
298 _hb_buffer_assert_unicode_vars (buffer);
299
300 hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference;
301 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_AUTO)
302 {
303 if (plan->has_gpos_mark)
304 // https://github.com/harfbuzz/harfbuzz/issues/653#issuecomment-423905920
305 //mode = HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED;
306 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
307 else
308 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
309 }
310
311 const hb_ot_shape_normalize_context_t c = {
312 plan,
313 buffer,
314 font,
315 buffer->unicode,
316 buffer->not_found,
317 plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode,
318 plan->shaper->compose ? plan->shaper->compose : compose_unicode
319 };
320
321 bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE;
322 bool might_short_circuit = always_short_circuit ||
323 (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED &&
324 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT);
325 unsigned int count;
326
327 /* We do a fairly straightforward yet custom normalization process in three
328 * separate rounds: decompose, reorder, recompose (if desired). Currently
329 * this makes two buffer swaps. We can make it faster by moving the last
330 * two rounds into the inner loop for the first round, but it's more readable
331 * this way. */
332
333
334 /* First round, decompose */
335
336 bool all_simple = true;
337 {
338 buffer->clear_output ();
339 count = buffer->len;
340 buffer->idx = 0;
341 do
342 {
343 unsigned int end;
344 for (end = buffer->idx + 1; end < count; end++)
345 if (unlikely (_hb_glyph_info_is_unicode_mark (&buffer->info[end])))
346 break;
347
348 if (end < count)
349 end--; /* Leave one base for the marks to cluster with. */
350
351 /* From idx to end are simple clusters. */
352 if (might_short_circuit)
353 {
354 unsigned int done = font->get_nominal_glyphs (end - buffer->idx,
355 &buffer->cur().codepoint,
356 sizeof (buffer->info[0]),
357 &buffer->cur().glyph_index(),
358 sizeof (buffer->info[0]));
359 if (unlikely (!buffer->next_glyphs (done))) break;
360 }
361 while (buffer->idx < end && buffer->successful)
362 decompose_current_character (&c, might_short_circuit);
363
364 if (buffer->idx == count || !buffer->successful)
365 break;
366
367 all_simple = false;
368
369 /* Find all the marks now. */
370 for (end = buffer->idx + 1; end < count; end++)
371 if (!_hb_glyph_info_is_unicode_mark(&buffer->info[end]))
372 break;
373
374 /* idx to end is one non-simple cluster. */
375 decompose_multi_char_cluster (&c, end, always_short_circuit);
376 }
377 while (buffer->idx < count && buffer->successful);
378 buffer->sync ();
379 }
380
381
382 /* Second round, reorder (inplace) */
383
384 if (!all_simple && buffer->message(font, "start reorder"))
385 {
386 count = buffer->len;
387 for (unsigned int i = 0; i < count; i++)
388 {
389 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0)
390 continue;
391
392 unsigned int end;
393 for (end = i + 1; end < count; end++)
394 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0)
395 break;
396
397 /* We are going to do a O(n^2). Only do this if the sequence is short. */
398 if (end - i > HB_OT_SHAPE_COMPLEX_MAX_COMBINING_MARKS) {
399 i = end;
400 continue;
401 }
402
403 buffer->sort (i, end, compare_combining_class);
404
405 if (plan->shaper->reorder_marks)
406 plan->shaper->reorder_marks (plan, buffer, i, end);
407
408 i = end;
409 }
410 (void) buffer->message(font, "end reorder");
411 }
412 if (buffer->scratch_flags & HB_BUFFER_SCRATCH_FLAG_HAS_CGJ)
413 {
414 /* For all CGJ, check if it prevented any reordering at all.
415 * If it did NOT, then make it skippable.
416 * https://github.com/harfbuzz/harfbuzz/issues/554
417 */
418 for (unsigned int i = 1; i + 1 < buffer->len; i++)
419 if (buffer->info[i].codepoint == 0x034Fu/*CGJ*/ &&
420 (info_cc(buffer->info[i+1]) == 0 || info_cc(buffer->info[i-1]) <= info_cc(buffer->info[i+1])))
421 {
422 _hb_glyph_info_unhide (&buffer->info[i]);
423 }
424 }
425
426
427 /* Third round, recompose */
428
429 if (!all_simple &&
430 buffer->successful &&
431 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS ||
432 mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT))
433 {
434 /* As noted in the comment earlier, we don't try to combine
435 * ccc=0 chars with their previous Starter. */
436
437 buffer->clear_output ();
438 count = buffer->len;
439 unsigned int starter = 0;
440 (void) buffer->next_glyph ();
441 while (buffer->idx < count /* No need for: && buffer->successful */)
442 {
443 hb_codepoint_t composed, glyph;
444 if (/* We don't try to compose a non-mark character with it's preceding starter.
445 * This is both an optimization to avoid trying to compose every two neighboring
446 * glyphs in most scripts AND a desired feature for Hangul. Apparently Hangul
447 * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */
448 _hb_glyph_info_is_unicode_mark(&buffer->cur()))
449 {
450 if (/* If there's anything between the starter and this char, they should have CCC
451 * smaller than this character's. */
452 (starter == buffer->out_len - 1 ||
453 info_cc (buffer->prev()) < info_cc (buffer->cur())) &&
454 /* And compose. */
455 c.compose (&c,
456 buffer->out_info[starter].codepoint,
457 buffer->cur().codepoint,
458 &composed) &&
459 /* And the font has glyph for the composite. */
460 font->get_nominal_glyph (composed, &glyph))
461 {
462 /* Composes. */
463 if (unlikely (!buffer->next_glyph ())) break; /* Copy to out-buffer. */
464 buffer->merge_out_clusters (starter, buffer->out_len);
465 buffer->out_len--; /* Remove the second composable. */
466 /* Modify starter and carry on. */
467 buffer->out_info[starter].codepoint = composed;
468 buffer->out_info[starter].glyph_index() = glyph;
469 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer);
470
471 continue;
472 }
473 }
474
475 /* Blocked, or doesn't compose. */
476 if (unlikely (!buffer->next_glyph ())) break;
477
478 if (info_cc (buffer->prev()) == 0)
479 starter = buffer->out_len - 1;
480 }
481 buffer->sync ();
482 }
483 }
484
485
486 #endif
487