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.hh"
28#include "hb-ot-shape-complex.hh"
29#include "hb-ot-shape.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 canonical
66 * decomposition, well, use the decomposition.
67 *
68 * - The complex shapers can customize the compose and decompose functions to
69 * offload some of their requirements to the normalizer. For example, the
70 * Indic shaper may want to disallow recomposing of two matras.
71 */
72
73static bool
74decompose_unicode (const hb_ot_shape_normalize_context_t *c,
75 hb_codepoint_t ab,
76 hb_codepoint_t *a,
77 hb_codepoint_t *b)
78{
79 return (bool) c->unicode->decompose (ab, a, b);
80}
81
82static bool
83compose_unicode (const hb_ot_shape_normalize_context_t *c,
84 hb_codepoint_t a,
85 hb_codepoint_t b,
86 hb_codepoint_t *ab)
87{
88 return (bool) c->unicode->compose (a, b, ab);
89}
90
91static inline void
92set_glyph (hb_glyph_info_t &info, hb_font_t *font)
93{
94 (void) font->get_nominal_glyph (info.codepoint, &info.glyph_index());
95}
96
97static inline void
98output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph)
99{
100 buffer->cur().glyph_index() = glyph;
101 buffer->output_glyph (unichar); /* This is very confusing indeed. */
102 _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer);
103}
104
105static inline void
106next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)
107{
108 buffer->cur().glyph_index() = glyph;
109 buffer->next_glyph ();
110}
111
112static inline void
113skip_char (hb_buffer_t *buffer)
114{
115 buffer->skip_glyph ();
116}
117
118/* Returns 0 if didn't decompose, number of resulting characters otherwise. */
119static inline unsigned int
120decompose (const hb_ot_shape_normalize_context_t *c, bool shortest, hb_codepoint_t ab)
121{
122 hb_codepoint_t a = 0, b = 0, a_glyph = 0, b_glyph = 0;
123 hb_buffer_t * const buffer = c->buffer;
124 hb_font_t * const font = c->font;
125
126 if (!c->decompose (c, ab, &a, &b) ||
127 (b && !font->get_nominal_glyph (b, &b_glyph)))
128 return 0;
129
130 bool has_a = (bool) font->get_nominal_glyph (a, &a_glyph);
131 if (shortest && has_a) {
132 /* Output a and b */
133 output_char (buffer, a, a_glyph);
134 if (likely (b)) {
135 output_char (buffer, b, b_glyph);
136 return 2;
137 }
138 return 1;
139 }
140
141 unsigned int ret;
142 if ((ret = decompose (c, shortest, a))) {
143 if (b) {
144 output_char (buffer, b, b_glyph);
145 return ret + 1;
146 }
147 return ret;
148 }
149
150 if (has_a) {
151 output_char (buffer, a, a_glyph);
152 if (likely (b)) {
153 output_char (buffer, b, b_glyph);
154 return 2;
155 }
156 return 1;
157 }
158
159 return 0;
160}
161
162static inline void
163decompose_current_character (const hb_ot_shape_normalize_context_t *c, bool shortest)
164{
165 hb_buffer_t * const buffer = c->buffer;
166 hb_codepoint_t u = buffer->cur().codepoint;
167 hb_codepoint_t glyph = 0;
168
169 if (shortest && c->font->get_nominal_glyph (u, &glyph))
170 {
171 next_char (buffer, glyph);
172 return;
173 }
174
175 if (decompose (c, shortest, u))
176 {
177 skip_char (buffer);
178 return;
179 }
180
181 if (!shortest && c->font->get_nominal_glyph (u, &glyph))
182 {
183 next_char (buffer, glyph);
184 return;
185 }
186
187 if (_hb_glyph_info_is_unicode_space (&buffer->cur()))
188 {
189 hb_codepoint_t space_glyph;
190 hb_unicode_funcs_t::space_t space_type = buffer->unicode->space_fallback_type (u);
191 if (space_type != hb_unicode_funcs_t::NOT_SPACE && c->font->get_nominal_glyph (0x0020u, &space_glyph))
192 {
193 _hb_glyph_info_set_unicode_space_fallback_type (&buffer->cur(), space_type);
194 next_char (buffer, space_glyph);
195 buffer->scratch_flags |= HB_BUFFER_SCRATCH_FLAG_HAS_SPACE_FALLBACK;
196 return;
197 }
198 }
199
200 if (u == 0x2011u)
201 {
202 /* U+2011 is the only sensible character that is a no-break version of another character
203 * and not a space. The space ones are handled already. Handle this lone one. */
204 hb_codepoint_t other_glyph;
205 if (c->font->get_nominal_glyph (0x2010u, &other_glyph))
206 {
207 next_char (buffer, other_glyph);
208 return;
209 }
210 }
211
212 next_char (buffer, glyph); /* glyph is initialized in earlier branches. */
213}
214
215static inline void
216handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c,
217 unsigned int end,
218 bool short_circuit HB_UNUSED)
219{
220 /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */
221 hb_buffer_t * const buffer = c->buffer;
222 hb_font_t * const font = c->font;
223 for (; buffer->idx < end - 1 && buffer->successful;) {
224 if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {
225 if (font->get_variation_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()))
226 {
227 hb_codepoint_t unicode = buffer->cur().codepoint;
228 buffer->replace_glyphs (2, 1, &unicode);
229 }
230 else
231 {
232 /* Just pass on the two characters separately, let GSUB do its magic. */
233 set_glyph (buffer->cur(), font);
234 buffer->next_glyph ();
235 set_glyph (buffer->cur(), font);
236 buffer->next_glyph ();
237 }
238 /* Skip any further variation selectors. */
239 while (buffer->idx < end && unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint)))
240 {
241 set_glyph (buffer->cur(), font);
242 buffer->next_glyph ();
243 }
244 } else {
245 set_glyph (buffer->cur(), font);
246 buffer->next_glyph ();
247 }
248 }
249 if (likely (buffer->idx < end)) {
250 set_glyph (buffer->cur(), font);
251 buffer->next_glyph ();
252 }
253}
254
255static inline void
256decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)
257{
258 hb_buffer_t * const buffer = c->buffer;
259 for (unsigned int i = buffer->idx; i < end && buffer->successful; i++)
260 if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {
261 handle_variation_selector_cluster (c, end, short_circuit);
262 return;
263 }
264
265 while (buffer->idx < end && buffer->successful)
266 decompose_current_character (c, short_circuit);
267}
268
269
270static int
271compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
272{
273 unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);
274 unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);
275
276 return a < b ? -1 : a == b ? 0 : +1;
277}
278
279
280void
281_hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,
282 hb_buffer_t *buffer,
283 hb_font_t *font)
284{
285 if (unlikely (!buffer->len)) return;
286
287 _hb_buffer_assert_unicode_vars (buffer);
288
289 hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference;
290 if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_AUTO)
291 {
292 if (plan->has_gpos_mark)
293 // https://github.com/harfbuzz/harfbuzz/issues/653#issuecomment-423905920
294 //mode = HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED;
295 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
296 else
297 mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
298 }
299
300 const hb_ot_shape_normalize_context_t c = {
301 plan,
302 buffer,
303 font,
304 buffer->unicode,
305 plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode,
306 plan->shaper->compose ? plan->shaper->compose : compose_unicode
307 };
308
309 bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE;
310 bool might_short_circuit = always_short_circuit ||
311 (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED &&
312 mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT);
313 unsigned int count;
314
315 /* We do a fairly straightforward yet custom normalization process in three
316 * separate rounds: decompose, reorder, recompose (if desired). Currently
317 * this makes two buffer swaps. We can make it faster by moving the last
318 * two rounds into the inner loop for the first round, but it's more readable
319 * this way. */
320
321
322 /* First round, decompose */
323
324 bool all_simple = true;
325 {
326 buffer->clear_output ();
327 count = buffer->len;
328 buffer->idx = 0;
329 do
330 {
331 unsigned int end;
332 for (end = buffer->idx + 1; end < count; end++)
333 if (unlikely (HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->info[end]))))
334 break;
335
336 if (end < count)
337 end--; /* Leave one base for the marks to cluster with. */
338
339 /* From idx to end are simple clusters. */
340 if (might_short_circuit)
341 {
342 unsigned int done = font->get_nominal_glyphs (end - buffer->idx,
343 &buffer->cur().codepoint,
344 sizeof (buffer->info[0]),
345 &buffer->cur().glyph_index(),
346 sizeof (buffer->info[0]));
347 buffer->next_glyphs (done);
348 }
349 while (buffer->idx < end && buffer->successful)
350 decompose_current_character (&c, might_short_circuit);
351
352 if (buffer->idx == count || !buffer->successful)
353 break;
354
355 all_simple = false;
356
357 /* Find all the marks now. */
358 for (end = buffer->idx + 1; end < count; end++)
359 if (!HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->info[end])))
360 break;
361
362 /* idx to end is one non-simple cluster. */
363 decompose_multi_char_cluster (&c, end, always_short_circuit);
364 }
365 while (buffer->idx < count && buffer->successful);
366 buffer->swap_buffers ();
367 }
368
369
370 /* Second round, reorder (inplace) */
371
372 if (!all_simple)
373 {
374 count = buffer->len;
375 for (unsigned int i = 0; i < count; i++)
376 {
377 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0)
378 continue;
379
380 unsigned int end;
381 for (end = i + 1; end < count; end++)
382 if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0)
383 break;
384
385 /* We are going to do a O(n^2). Only do this if the sequence is short. */
386 if (end - i > HB_OT_SHAPE_COMPLEX_MAX_COMBINING_MARKS) {
387 i = end;
388 continue;
389 }
390
391 buffer->sort (i, end, compare_combining_class);
392
393 if (plan->shaper->reorder_marks)
394 plan->shaper->reorder_marks (plan, buffer, i, end);
395
396 i = end;
397 }
398 }
399 if (buffer->scratch_flags & HB_BUFFER_SCRATCH_FLAG_HAS_CGJ)
400 {
401 /* For all CGJ, check if it prevented any reordering at all.
402 * If it did NOT, then make it skippable.
403 * https://github.com/harfbuzz/harfbuzz/issues/554
404 */
405 for (unsigned int i = 1; i + 1 < buffer->len; i++)
406 if (buffer->info[i].codepoint == 0x034Fu/*CGJ*/ &&
407 info_cc(buffer->info[i-1]) <= info_cc(buffer->info[i+1]))
408 {
409 _hb_glyph_info_unhide (&buffer->info[i]);
410 }
411 }
412
413
414 /* Third round, recompose */
415
416 if (!all_simple &&
417 (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS ||
418 mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT))
419 {
420 /* As noted in the comment earlier, we don't try to combine
421 * ccc=0 chars with their previous Starter. */
422
423 buffer->clear_output ();
424 count = buffer->len;
425 unsigned int starter = 0;
426 buffer->next_glyph ();
427 while (buffer->idx < count && buffer->successful)
428 {
429 hb_codepoint_t composed, glyph;
430 if (/* We don't try to compose a non-mark character with it's preceding starter.
431 * This is both an optimization to avoid trying to compose every two neighboring
432 * glyphs in most scripts AND a desired feature for Hangul. Apparently Hangul
433 * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */
434 HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->cur())))
435 {
436 if (/* If there's anything between the starter and this char, they should have CCC
437 * smaller than this character's. */
438 (starter == buffer->out_len - 1 ||
439 info_cc (buffer->prev()) < info_cc (buffer->cur())) &&
440 /* And compose. */
441 c.compose (&c,
442 buffer->out_info[starter].codepoint,
443 buffer->cur().codepoint,
444 &composed) &&
445 /* And the font has glyph for the composite. */
446 font->get_nominal_glyph (composed, &glyph))
447 {
448 /* Composes. */
449 buffer->next_glyph (); /* Copy to out-buffer. */
450 if (unlikely (!buffer->successful))
451 return;
452 buffer->merge_out_clusters (starter, buffer->out_len);
453 buffer->out_len--; /* Remove the second composable. */
454 /* Modify starter and carry on. */
455 buffer->out_info[starter].codepoint = composed;
456 buffer->out_info[starter].glyph_index() = glyph;
457 _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer);
458
459 continue;
460 }
461 }
462
463 /* Blocked, or doesn't compose. */
464 buffer->next_glyph ();
465
466 if (info_cc (buffer->prev()) == 0)
467 starter = buffer->out_len - 1;
468 }
469 buffer->swap_buffers ();
470 }
471}
472