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-shaper.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 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
77static bool
78decompose_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
86static bool
87compose_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
95static inline void
96set_glyph (hb_glyph_info_t &info, hb_font_t *font)
97{
98 (void) font->get_nominal_glyph (info.codepoint, &info.glyph_index());
99}
100
101static inline void
102output_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
110static inline void
111next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)
112{
113 buffer->cur().glyph_index() = glyph;
114 (void) buffer->next_glyph ();
115}
116
117static inline void
118skip_char (hb_buffer_t *buffer)
119{
120 buffer->skip_glyph ();
121}
122
123/* Returns 0 if didn't decompose, number of resulting characters otherwise. */
124static inline unsigned int
125decompose (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 if (unsigned ret = decompose (c, shortest, a)) {
147 if (b) {
148 output_char (buffer, b, b_glyph);
149 return ret + 1;
150 }
151 return ret;
152 }
153
154 if (has_a) {
155 output_char (buffer, a, a_glyph);
156 if (likely (b)) {
157 output_char (buffer, b, b_glyph);
158 return 2;
159 }
160 return 1;
161 }
162
163 return 0;
164}
165
166static inline void
167decompose_current_character (const hb_ot_shape_normalize_context_t *c, bool shortest)
168{
169 hb_buffer_t * const buffer = c->buffer;
170 hb_codepoint_t u = buffer->cur().codepoint;
171 hb_codepoint_t glyph = 0;
172
173 if (shortest && c->font->get_nominal_glyph (u, &glyph, c->not_found))
174 {
175 next_char (buffer, glyph);
176 return;
177 }
178
179 if (decompose (c, shortest, u))
180 {
181 skip_char (buffer);
182 return;
183 }
184
185 if (!shortest && c->font->get_nominal_glyph (u, &glyph, c->not_found))
186 {
187 next_char (buffer, glyph);
188 return;
189 }
190
191 if (_hb_glyph_info_is_unicode_space (&buffer->cur()))
192 {
193 hb_codepoint_t space_glyph;
194 hb_unicode_funcs_t::space_t space_type = buffer->unicode->space_fallback_type (u);
195 if (space_type != hb_unicode_funcs_t::NOT_SPACE &&
196 (c->font->get_nominal_glyph (0x0020, &space_glyph) || (space_glyph = buffer->invisible)))
197 {
198 _hb_glyph_info_set_unicode_space_fallback_type (&buffer->cur(), space_type);
199 next_char (buffer, space_glyph);
200 buffer->scratch_flags |= HB_BUFFER_SCRATCH_FLAG_HAS_SPACE_FALLBACK;
201 return;
202 }
203 }
204
205 if (u == 0x2011u)
206 {
207 /* U+2011 is the only sensible character that is a no-break version of another character
208 * and not a space. The space ones are handled already. Handle this lone one. */
209 hb_codepoint_t other_glyph;
210 if (c->font->get_nominal_glyph (0x2010u, &other_glyph))
211 {
212 next_char (buffer, other_glyph);
213 return;
214 }
215 }
216
217 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */
218}
219
220static inline void
221handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c,
222 unsigned int end,
223 bool short_circuit HB_UNUSED)
224{
225 /* Currently if there's a variation-selector we give-up on normalization, it's just too hard. */
226 hb_buffer_t * const buffer = c->buffer;
227 hb_font_t * const font = c->font;
228 for (; buffer->idx < end - 1 && buffer->successful;) {
229 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {
230 if (font->get_variation_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()))
231 {
232 hb_codepoint_t unicode = buffer->cur().codepoint;
233 (void) buffer->replace_glyphs (2, 1, &unicode);
234 }
235 else
236 {
237 /* Just pass on the two characters separately, let GSUB do its magic. */
238 set_glyph (buffer->cur(), font);
239 (void) buffer->next_glyph ();
240 set_glyph (buffer->cur(), font);
241 (void) buffer->next_glyph ();
242 }
243 /* Skip any further variation selectors. */
244 while (buffer->idx < end &&
245 buffer->successful &&
246 unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint)))
247 {
248 set_glyph (buffer->cur(), font);
249 (void) buffer->next_glyph ();
250 }
251 }
252 else
253 {
254 set_glyph (buffer->cur(), font);
255 (void) buffer->next_glyph ();
256 }
257 }
258 if (likely (buffer->idx < end))
259 {
260 set_glyph (buffer->cur(), font);
261 (void) buffer->next_glyph ();
262 }
263}
264
265static inline void
266decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)
267{
268 hb_buffer_t * const buffer = c->buffer;
269 for (unsigned int i = buffer->idx; i < end && buffer->successful; i++)
270 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {
271 handle_variation_selector_cluster (c, end, short_circuit);
272 return;
273 }
274
275 while (buffer->idx < end && buffer->successful)
276 decompose_current_character (c, short_circuit);
277}
278
279
280static int
281compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
282{
283 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);
284 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);
285
286 return a < b ? -1 : a == b ? 0 : +1;
287}
288
289
290void
291_hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,
292 hb_buffer_t *buffer,
293 hb_font_t *font)
294{
295 if (unlikely (!buffer->len)) return;
296
297 _hb_buffer_assert_unicode_vars (buffer);
298
299 hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference;
300 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_AUTO)
301 {
302 if (plan->has_gpos_mark)
303 // https://github.com/harfbuzz/harfbuzz/issues/653#issuecomment-423905920
304 //mode = HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED;
305 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
306 else
307 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
308 }
309
310 const hb_ot_shape_normalize_context_t c = {
311 plan,
312 buffer,
313 font,
314 buffer->unicode,
315 buffer->not_found,
316 plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode,
317 plan->shaper->compose ? plan->shaper->compose : compose_unicode
318 };
319
320 bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE;
321 bool might_short_circuit = always_short_circuit ||
322 (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED &&
323 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT);
324 unsigned int count;
325
326 /* We do a fairly straightforward yet custom normalization process in three
327 * separate rounds: decompose, reorder, recompose (if desired). Currently
328 * this makes two buffer swaps. We can make it faster by moving the last
329 * two rounds into the inner loop for the first round, but it's more readable
330 * this way. */
331
332
333 /* First round, decompose */
334
335 bool all_simple = true;
336 {
337 buffer->clear_output ();
338 count = buffer->len;
339 buffer->idx = 0;
340 do
341 {
342 unsigned int end;
343 for (end = buffer->idx + 1; end < count; end++)
344 if (_hb_glyph_info_is_unicode_mark (&buffer->info[end]))
345 break;
346
347 if (end < count)
348 end--; /* Leave one base for the marks to cluster with. */
349
350 /* From idx to end are simple clusters. */
351 if (might_short_circuit)
352 {
353 unsigned int done = font->get_nominal_glyphs (end - buffer->idx,
354 &buffer->cur().codepoint,
355 sizeof (buffer->info[0]),
356 &buffer->cur().glyph_index(),
357 sizeof (buffer->info[0]));
358 if (unlikely (!buffer->next_glyphs (done))) break;
359 }
360 while (buffer->idx < end && buffer->successful)
361 decompose_current_character (&c, might_short_circuit);
362
363 if (buffer->idx == count || !buffer->successful)
364 break;
365
366 all_simple = false;
367
368 /* Find all the marks now. */
369 for (end = buffer->idx + 1; end < count; end++)
370 if (!_hb_glyph_info_is_unicode_mark(&buffer->info[end]))
371 break;
372
373 /* idx to end is one non-simple cluster. */
374 decompose_multi_char_cluster (&c, end, always_short_circuit);
375 }
376 while (buffer->idx < count && buffer->successful);
377 buffer->sync ();
378 }
379
380
381 /* Second round, reorder (inplace) */
382
383 if (!all_simple && buffer->message(font, "start reorder"))
384 {
385 count = buffer->len;
386 hb_glyph_info_t *info = buffer->info;
387 for (unsigned int i = 0; i < count; i++)
388 {
389 if (_hb_glyph_info_get_modified_combining_class (&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 (&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_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 unsigned count = buffer->len;
419 hb_glyph_info_t *info = buffer->info;
420 for (unsigned int i = 1; i + 1 < count; i++)
421 if (info[i].codepoint == 0x034Fu/*CGJ*/ &&
422 (info_cc(info[i+1]) == 0 || info_cc(info[i-1]) <= info_cc(info[i+1])))
423 {
424 _hb_glyph_info_unhide (&info[i]);
425 }
426 }
427
428
429 /* Third round, recompose */
430
431 if (!all_simple &&
432 buffer->successful &&
433 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS ||
434 mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT))
435 {
436 /* As noted in the comment earlier, we don't try to combine
437 * ccc=0 chars with their previous Starter. */
438
439 buffer->clear_output ();
440 count = buffer->len;
441 unsigned int starter = 0;
442 (void) buffer->next_glyph ();
443 while (buffer->idx < count /* No need for: && buffer->successful */)
444 {
445 hb_codepoint_t composed, glyph;
446 if (/* We don't try to compose a non-mark character with it's preceding starter.
447 * This is both an optimization to avoid trying to compose every two neighboring
448 * glyphs in most scripts AND a desired feature for Hangul. Apparently Hangul
449 * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */
450 _hb_glyph_info_is_unicode_mark(&buffer->cur()))
451 {
452 if (/* If there's anything between the starter and this char, they should have CCC
453 * smaller than this character's. */
454 (starter == buffer->out_len - 1 ||
455 info_cc (buffer->prev()) < info_cc (buffer->cur())) &&
456 /* And compose. */
457 c.compose (&c,
458 buffer->out_info[starter].codepoint,
459 buffer->cur().codepoint,
460 &composed) &&
461 /* And the font has glyph for the composite. */
462 font->get_nominal_glyph (composed, &glyph))
463 {
464 /* Composes. */
465 if (unlikely (!buffer->next_glyph ())) break; /* Copy to out-buffer. */
466 buffer->merge_out_clusters (starter, buffer->out_len);
467 buffer->out_len--; /* Remove the second composable. */
468 /* Modify starter and carry on. */
469 buffer->out_info[starter].codepoint = composed;
470 buffer->out_info[starter].glyph_index() = glyph;
471 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer);
472
473 continue;
474 }
475 }
476
477 /* Blocked, or doesn't compose. */
478 if (unlikely (!buffer->next_glyph ())) break;
479
480 if (info_cc (buffer->prev()) == 0)
481 starter = buffer->out_len - 1;
482 }
483 buffer->sync ();
484 }
485}
486
487
488#endif
489