• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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-complex-indic.hh"
28 #include "hb-ot-shape-complex-vowel-constraints.hh"
29 #include "hb-ot-layout.hh"
30 
31 
32 /*
33  * Indic shaper.
34  */
35 
36 
37 /*
38  * Indic configurations.  Note that we do not want to keep every single script-specific
39  * behavior in these tables necessarily.  This should mainly be used for per-script
40  * properties that are cheaper keeping here, than in the code.  Ie. if, say, one and
41  * only one script has an exception, that one script can be if'ed directly in the code,
42  * instead of adding a new flag in these structs.
43  */
44 
45 enum base_position_t {
46   BASE_POS_LAST_SINHALA,
47   BASE_POS_LAST
48 };
49 enum reph_position_t {
50   REPH_POS_AFTER_MAIN  = POS_AFTER_MAIN,
51   REPH_POS_BEFORE_SUB  = POS_BEFORE_SUB,
52   REPH_POS_AFTER_SUB   = POS_AFTER_SUB,
53   REPH_POS_BEFORE_POST = POS_BEFORE_POST,
54   REPH_POS_AFTER_POST  = POS_AFTER_POST
55 };
56 enum reph_mode_t {
57   REPH_MODE_IMPLICIT,  /* Reph formed out of initial Ra,H sequence. */
58   REPH_MODE_EXPLICIT,  /* Reph formed out of initial Ra,H,ZWJ sequence. */
59   REPH_MODE_LOG_REPHA  /* Encoded Repha character, needs reordering. */
60 };
61 enum blwf_mode_t {
62   BLWF_MODE_PRE_AND_POST, /* Below-forms feature applied to pre-base and post-base. */
63   BLWF_MODE_POST_ONLY     /* Below-forms feature applied to post-base only. */
64 };
65 struct indic_config_t
66 {
67   hb_script_t     script;
68   bool            has_old_spec;
69   hb_codepoint_t  virama;
70   base_position_t base_pos;
71   reph_position_t reph_pos;
72   reph_mode_t     reph_mode;
73   blwf_mode_t     blwf_mode;
74 };
75 
76 static const indic_config_t indic_configs[] =
77 {
78   /* Default.  Should be first. */
79   {HB_SCRIPT_INVALID,	false,      0,BASE_POS_LAST, REPH_POS_BEFORE_POST,REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
80   {HB_SCRIPT_DEVANAGARI,true, 0x094Du,BASE_POS_LAST, REPH_POS_BEFORE_POST,REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
81   {HB_SCRIPT_BENGALI,	true, 0x09CDu,BASE_POS_LAST, REPH_POS_AFTER_SUB,  REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
82   {HB_SCRIPT_GURMUKHI,	true, 0x0A4Du,BASE_POS_LAST, REPH_POS_BEFORE_SUB, REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
83   {HB_SCRIPT_GUJARATI,	true, 0x0ACDu,BASE_POS_LAST, REPH_POS_BEFORE_POST,REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
84   {HB_SCRIPT_ORIYA,	true, 0x0B4Du,BASE_POS_LAST, REPH_POS_AFTER_MAIN, REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
85   {HB_SCRIPT_TAMIL,	true, 0x0BCDu,BASE_POS_LAST, REPH_POS_AFTER_POST, REPH_MODE_IMPLICIT, BLWF_MODE_PRE_AND_POST},
86   {HB_SCRIPT_TELUGU,	true, 0x0C4Du,BASE_POS_LAST, REPH_POS_AFTER_POST, REPH_MODE_EXPLICIT, BLWF_MODE_POST_ONLY},
87   {HB_SCRIPT_KANNADA,	true, 0x0CCDu,BASE_POS_LAST, REPH_POS_AFTER_POST, REPH_MODE_IMPLICIT, BLWF_MODE_POST_ONLY},
88   {HB_SCRIPT_MALAYALAM,	true, 0x0D4Du,BASE_POS_LAST, REPH_POS_AFTER_MAIN, REPH_MODE_LOG_REPHA,BLWF_MODE_PRE_AND_POST},
89   {HB_SCRIPT_SINHALA,	false,0x0DCAu,BASE_POS_LAST_SINHALA,
90 						     REPH_POS_AFTER_POST, REPH_MODE_EXPLICIT, BLWF_MODE_PRE_AND_POST},
91 };
92 
93 
94 
95 /*
96  * Indic shaper.
97  */
98 
99 static const hb_ot_map_feature_t
100 indic_features[] =
101 {
102   /*
103    * Basic features.
104    * These features are applied in order, one at a time, after initial_reordering.
105    */
106   {HB_TAG('n','u','k','t'), F_GLOBAL_MANUAL_JOINERS},
107   {HB_TAG('a','k','h','n'), F_GLOBAL_MANUAL_JOINERS},
108   {HB_TAG('r','p','h','f'),        F_MANUAL_JOINERS},
109   {HB_TAG('r','k','r','f'), F_GLOBAL_MANUAL_JOINERS},
110   {HB_TAG('p','r','e','f'),        F_MANUAL_JOINERS},
111   {HB_TAG('b','l','w','f'),        F_MANUAL_JOINERS},
112   {HB_TAG('a','b','v','f'),        F_MANUAL_JOINERS},
113   {HB_TAG('h','a','l','f'),        F_MANUAL_JOINERS},
114   {HB_TAG('p','s','t','f'),        F_MANUAL_JOINERS},
115   {HB_TAG('v','a','t','u'), F_GLOBAL_MANUAL_JOINERS},
116   {HB_TAG('c','j','c','t'), F_GLOBAL_MANUAL_JOINERS},
117   /*
118    * Other features.
119    * These features are applied all at once, after final_reordering
120    * but before clearing syllables.
121    * Default Bengali font in Windows for example has intermixed
122    * lookups for init,pres,abvs,blws features.
123    */
124   {HB_TAG('i','n','i','t'),        F_MANUAL_JOINERS},
125   {HB_TAG('p','r','e','s'), F_GLOBAL_MANUAL_JOINERS},
126   {HB_TAG('a','b','v','s'), F_GLOBAL_MANUAL_JOINERS},
127   {HB_TAG('b','l','w','s'), F_GLOBAL_MANUAL_JOINERS},
128   {HB_TAG('p','s','t','s'), F_GLOBAL_MANUAL_JOINERS},
129   {HB_TAG('h','a','l','n'), F_GLOBAL_MANUAL_JOINERS},
130   /*
131    * Positioning features.
132    * We don't care about the types.
133    */
134   {HB_TAG('d','i','s','t'), F_GLOBAL},
135   {HB_TAG('a','b','v','m'), F_GLOBAL},
136   {HB_TAG('b','l','w','m'), F_GLOBAL},
137 };
138 
139 /*
140  * Must be in the same order as the indic_features array.
141  */
142 enum {
143   _NUKT,
144   _AKHN,
145   RPHF,
146   _RKRF,
147   PREF,
148   BLWF,
149   ABVF,
150   HALF,
151   PSTF,
152   _VATU,
153   _CJCT,
154 
155   INIT,
156   _PRES,
157   _ABVS,
158   _BLWS,
159   _PSTS,
160   _HALN,
161 
162   _DIST,
163   _ABVM,
164   _BLWM,
165 
166   INDIC_NUM_FEATURES,
167   INDIC_BASIC_FEATURES = INIT, /* Don't forget to update this! */
168 };
169 
170 static void
171 setup_syllables (const hb_ot_shape_plan_t *plan,
172 		 hb_font_t *font,
173 		 hb_buffer_t *buffer);
174 static void
175 initial_reordering (const hb_ot_shape_plan_t *plan,
176 		    hb_font_t *font,
177 		    hb_buffer_t *buffer);
178 static void
179 final_reordering (const hb_ot_shape_plan_t *plan,
180 		  hb_font_t *font,
181 		  hb_buffer_t *buffer);
182 static void
183 clear_syllables (const hb_ot_shape_plan_t *plan,
184 		 hb_font_t *font,
185 		 hb_buffer_t *buffer);
186 
187 static void
collect_features_indic(hb_ot_shape_planner_t * plan)188 collect_features_indic (hb_ot_shape_planner_t *plan)
189 {
190   hb_ot_map_builder_t *map = &plan->map;
191 
192   /* Do this before any lookups have been applied. */
193   map->add_gsub_pause (setup_syllables);
194 
195   map->enable_feature (HB_TAG('l','o','c','l'));
196   /* The Indic specs do not require ccmp, but we apply it here since if
197    * there is a use of it, it's typically at the beginning. */
198   map->enable_feature (HB_TAG('c','c','m','p'));
199 
200 
201   unsigned int i = 0;
202   map->add_gsub_pause (initial_reordering);
203 
204   for (; i < INDIC_BASIC_FEATURES; i++) {
205     map->add_feature (indic_features[i]);
206     map->add_gsub_pause (nullptr);
207   }
208 
209   map->add_gsub_pause (final_reordering);
210 
211   for (; i < INDIC_NUM_FEATURES; i++)
212     map->add_feature (indic_features[i]);
213 
214   map->enable_feature (HB_TAG('c','a','l','t'));
215   map->enable_feature (HB_TAG('c','l','i','g'));
216 
217   map->add_gsub_pause (clear_syllables);
218 }
219 
220 static void
override_features_indic(hb_ot_shape_planner_t * plan)221 override_features_indic (hb_ot_shape_planner_t *plan)
222 {
223   plan->map.disable_feature (HB_TAG('l','i','g','a'));
224 }
225 
226 
227 struct would_substitute_feature_t
228 {
initwould_substitute_feature_t229   void init (const hb_ot_map_t *map, hb_tag_t feature_tag, bool zero_context_)
230   {
231     zero_context = zero_context_;
232     map->get_stage_lookups (0/*GSUB*/,
233 			    map->get_feature_stage (0/*GSUB*/, feature_tag),
234 			    &lookups, &count);
235   }
236 
would_substitutewould_substitute_feature_t237   bool would_substitute (const hb_codepoint_t *glyphs,
238 			 unsigned int          glyphs_count,
239 			 hb_face_t            *face) const
240   {
241     for (unsigned int i = 0; i < count; i++)
242       if (hb_ot_layout_lookup_would_substitute_fast (face, lookups[i].index, glyphs, glyphs_count, zero_context))
243 	return true;
244     return false;
245   }
246 
247   private:
248   const hb_ot_map_t::lookup_map_t *lookups;
249   unsigned int count;
250   bool zero_context;
251 };
252 
253 struct indic_shape_plan_t
254 {
load_virama_glyphindic_shape_plan_t255   bool load_virama_glyph (hb_font_t *font, hb_codepoint_t *pglyph) const
256   {
257     hb_codepoint_t glyph = virama_glyph.get_relaxed ();
258     if (unlikely (glyph == (hb_codepoint_t) -1))
259     {
260       if (!config->virama || !font->get_nominal_glyph (config->virama, &glyph))
261 	glyph = 0;
262       /* Technically speaking, the spec says we should apply 'locl' to virama too.
263        * Maybe one day... */
264 
265       /* Our get_nominal_glyph() function needs a font, so we can't get the virama glyph
266        * during shape planning...  Instead, overwrite it here. */
267       virama_glyph.set_relaxed ((int) glyph);
268     }
269 
270     *pglyph = glyph;
271     return glyph != 0;
272   }
273 
274   const indic_config_t *config;
275 
276   bool is_old_spec;
277   bool uniscribe_bug_compatible;
278   mutable hb_atomic_int_t virama_glyph;
279 
280   would_substitute_feature_t rphf;
281   would_substitute_feature_t pref;
282   would_substitute_feature_t blwf;
283   would_substitute_feature_t pstf;
284 
285   hb_mask_t mask_array[INDIC_NUM_FEATURES];
286 };
287 
288 static void *
data_create_indic(const hb_ot_shape_plan_t * plan)289 data_create_indic (const hb_ot_shape_plan_t *plan)
290 {
291   indic_shape_plan_t *indic_plan = (indic_shape_plan_t *) calloc (1, sizeof (indic_shape_plan_t));
292   if (unlikely (!indic_plan))
293     return nullptr;
294 
295   indic_plan->config = &indic_configs[0];
296   for (unsigned int i = 1; i < ARRAY_LENGTH (indic_configs); i++)
297     if (plan->props.script == indic_configs[i].script) {
298       indic_plan->config = &indic_configs[i];
299       break;
300     }
301 
302   indic_plan->is_old_spec = indic_plan->config->has_old_spec && ((plan->map.chosen_script[0] & 0x000000FFu) != '2');
303   indic_plan->uniscribe_bug_compatible = hb_options ().uniscribe_bug_compatible;
304   indic_plan->virama_glyph.set_relaxed (-1);
305 
306   /* Use zero-context would_substitute() matching for new-spec of the main
307    * Indic scripts, and scripts with one spec only, but not for old-specs.
308    * The new-spec for all dual-spec scripts says zero-context matching happens.
309    *
310    * However, testing with Malayalam shows that old and new spec both allow
311    * context.  Testing with Bengali new-spec however shows that it doesn't.
312    * So, the heuristic here is the way it is.  It should *only* be changed,
313    * as we discover more cases of what Windows does.  DON'T TOUCH OTHERWISE.
314    */
315   bool zero_context = !indic_plan->is_old_spec && plan->props.script != HB_SCRIPT_MALAYALAM;
316   indic_plan->rphf.init (&plan->map, HB_TAG('r','p','h','f'), zero_context);
317   indic_plan->pref.init (&plan->map, HB_TAG('p','r','e','f'), zero_context);
318   indic_plan->blwf.init (&plan->map, HB_TAG('b','l','w','f'), zero_context);
319   indic_plan->pstf.init (&plan->map, HB_TAG('p','s','t','f'), zero_context);
320 
321   for (unsigned int i = 0; i < ARRAY_LENGTH (indic_plan->mask_array); i++)
322     indic_plan->mask_array[i] = (indic_features[i].flags & F_GLOBAL) ?
323 				 0 : plan->map.get_1_mask (indic_features[i].tag);
324 
325   return indic_plan;
326 }
327 
328 static void
data_destroy_indic(void * data)329 data_destroy_indic (void *data)
330 {
331   free (data);
332 }
333 
334 static indic_position_t
consonant_position_from_face(const indic_shape_plan_t * indic_plan,const hb_codepoint_t consonant,const hb_codepoint_t virama,hb_face_t * face)335 consonant_position_from_face (const indic_shape_plan_t *indic_plan,
336 			      const hb_codepoint_t consonant,
337 			      const hb_codepoint_t virama,
338 			      hb_face_t *face)
339 {
340   /* For old-spec, the order of glyphs is Consonant,Virama,
341    * whereas for new-spec, it's Virama,Consonant.  However,
342    * some broken fonts (like Free Sans) simply copied lookups
343    * from old-spec to new-spec without modification.
344    * And oddly enough, Uniscribe seems to respect those lookups.
345    * Eg. in the sequence U+0924,U+094D,U+0930, Uniscribe finds
346    * base at 0.  The font however, only has lookups matching
347    * 930,94D in 'blwf', not the expected 94D,930 (with new-spec
348    * table).  As such, we simply match both sequences.  Seems
349    * to work. */
350   hb_codepoint_t glyphs[3] = {virama, consonant, virama};
351   if (indic_plan->blwf.would_substitute (glyphs  , 2, face) ||
352       indic_plan->blwf.would_substitute (glyphs+1, 2, face))
353     return POS_BELOW_C;
354   if (indic_plan->pstf.would_substitute (glyphs  , 2, face) ||
355       indic_plan->pstf.would_substitute (glyphs+1, 2, face))
356     return POS_POST_C;
357   if (indic_plan->pref.would_substitute (glyphs  , 2, face) ||
358       indic_plan->pref.would_substitute (glyphs+1, 2, face))
359     return POS_POST_C;
360   return POS_BASE_C;
361 }
362 
363 
364 enum syllable_type_t {
365   consonant_syllable,
366   vowel_syllable,
367   standalone_cluster,
368   symbol_cluster,
369   broken_cluster,
370   non_indic_cluster,
371 };
372 
373 #include "hb-ot-shape-complex-indic-machine.hh"
374 
375 
376 static void
setup_masks_indic(const hb_ot_shape_plan_t * plan HB_UNUSED,hb_buffer_t * buffer,hb_font_t * font HB_UNUSED)377 setup_masks_indic (const hb_ot_shape_plan_t *plan HB_UNUSED,
378 		   hb_buffer_t              *buffer,
379 		   hb_font_t                *font HB_UNUSED)
380 {
381   HB_BUFFER_ALLOCATE_VAR (buffer, indic_category);
382   HB_BUFFER_ALLOCATE_VAR (buffer, indic_position);
383 
384   /* We cannot setup masks here.  We save information about characters
385    * and setup masks later on in a pause-callback. */
386 
387   unsigned int count = buffer->len;
388   hb_glyph_info_t *info = buffer->info;
389   for (unsigned int i = 0; i < count; i++)
390     set_indic_properties (info[i]);
391 }
392 
393 static void
setup_syllables(const hb_ot_shape_plan_t * plan HB_UNUSED,hb_font_t * font HB_UNUSED,hb_buffer_t * buffer)394 setup_syllables (const hb_ot_shape_plan_t *plan HB_UNUSED,
395 		 hb_font_t *font HB_UNUSED,
396 		 hb_buffer_t *buffer)
397 {
398   find_syllables (buffer);
399   foreach_syllable (buffer, start, end)
400     buffer->unsafe_to_break (start, end);
401 }
402 
403 static int
compare_indic_order(const hb_glyph_info_t * pa,const hb_glyph_info_t * pb)404 compare_indic_order (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
405 {
406   int a = pa->indic_position();
407   int b = pb->indic_position();
408 
409   return a < b ? -1 : a == b ? 0 : +1;
410 }
411 
412 
413 
414 static void
update_consonant_positions(const hb_ot_shape_plan_t * plan,hb_font_t * font,hb_buffer_t * buffer)415 update_consonant_positions (const hb_ot_shape_plan_t *plan,
416 			    hb_font_t         *font,
417 			    hb_buffer_t       *buffer)
418 {
419   const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) plan->data;
420 
421   if (indic_plan->config->base_pos != BASE_POS_LAST)
422     return;
423 
424   hb_codepoint_t virama;
425   if (indic_plan->load_virama_glyph (font, &virama))
426   {
427     hb_face_t *face = font->face;
428     unsigned int count = buffer->len;
429     hb_glyph_info_t *info = buffer->info;
430     for (unsigned int i = 0; i < count; i++)
431       if (info[i].indic_position() == POS_BASE_C)
432       {
433 	hb_codepoint_t consonant = info[i].codepoint;
434 	info[i].indic_position() = consonant_position_from_face (indic_plan, consonant, virama, face);
435       }
436   }
437 }
438 
439 
440 /* Rules from:
441  * https://docs.microsqoft.com/en-us/typography/script-development/devanagari */
442 
443 static void
initial_reordering_consonant_syllable(const hb_ot_shape_plan_t * plan,hb_face_t * face,hb_buffer_t * buffer,unsigned int start,unsigned int end)444 initial_reordering_consonant_syllable (const hb_ot_shape_plan_t *plan,
445 				       hb_face_t *face,
446 				       hb_buffer_t *buffer,
447 				       unsigned int start, unsigned int end)
448 {
449   const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) plan->data;
450   hb_glyph_info_t *info = buffer->info;
451 
452   /* https://github.com/harfbuzz/harfbuzz/issues/435#issuecomment-335560167
453    * // For compatibility with legacy usage in Kannada,
454    * // Ra+h+ZWJ must behave like Ra+ZWJ+h...
455    */
456   if (buffer->props.script == HB_SCRIPT_KANNADA &&
457       start + 3 <= end &&
458       is_one_of (info[start  ], FLAG (OT_Ra)) &&
459       is_one_of (info[start+1], FLAG (OT_H)) &&
460       is_one_of (info[start+2], FLAG (OT_ZWJ)))
461   {
462     buffer->merge_clusters (start+1, start+3);
463     hb_glyph_info_t tmp = info[start+1];
464     info[start+1] = info[start+2];
465     info[start+2] = tmp;
466   }
467 
468   /* 1. Find base consonant:
469    *
470    * The shaping engine finds the base consonant of the syllable, using the
471    * following algorithm: starting from the end of the syllable, move backwards
472    * until a consonant is found that does not have a below-base or post-base
473    * form (post-base forms have to follow below-base forms), or that is not a
474    * pre-base-reordering Ra, or arrive at the first consonant. The consonant
475    * stopped at will be the base.
476    *
477    *   o If the syllable starts with Ra + Halant (in a script that has Reph)
478    *     and has more than one consonant, Ra is excluded from candidates for
479    *     base consonants.
480    */
481 
482   unsigned int base = end;
483   bool has_reph = false;
484 
485   {
486     /* -> If the syllable starts with Ra + Halant (in a script that has Reph)
487      *    and has more than one consonant, Ra is excluded from candidates for
488      *    base consonants. */
489     unsigned int limit = start;
490     if (indic_plan->mask_array[RPHF] &&
491 	start + 3 <= end &&
492 	(
493 	 (indic_plan->config->reph_mode == REPH_MODE_IMPLICIT && !is_joiner (info[start + 2])) ||
494 	 (indic_plan->config->reph_mode == REPH_MODE_EXPLICIT && info[start + 2].indic_category() == OT_ZWJ)
495 	))
496     {
497       /* See if it matches the 'rphf' feature. */
498       hb_codepoint_t glyphs[3] = {info[start].codepoint,
499 				  info[start + 1].codepoint,
500 				  indic_plan->config->reph_mode == REPH_MODE_EXPLICIT ?
501 				    info[start + 2].codepoint : 0};
502       if (indic_plan->rphf.would_substitute (glyphs, 2, face) ||
503 	  (indic_plan->config->reph_mode == REPH_MODE_EXPLICIT &&
504 	   indic_plan->rphf.would_substitute (glyphs, 3, face)))
505       {
506 	limit += 2;
507 	while (limit < end && is_joiner (info[limit]))
508 	  limit++;
509 	base = start;
510 	has_reph = true;
511       }
512     } else if (indic_plan->config->reph_mode == REPH_MODE_LOG_REPHA && info[start].indic_category() == OT_Repha)
513     {
514 	limit += 1;
515 	while (limit < end && is_joiner (info[limit]))
516 	  limit++;
517 	base = start;
518 	has_reph = true;
519     }
520 
521     switch (indic_plan->config->base_pos)
522     {
523       case BASE_POS_LAST:
524       {
525 	/* -> starting from the end of the syllable, move backwards */
526 	unsigned int i = end;
527 	bool seen_below = false;
528 	do {
529 	  i--;
530 	  /* -> until a consonant is found */
531 	  if (is_consonant (info[i]))
532 	  {
533 	    /* -> that does not have a below-base or post-base form
534 	     * (post-base forms have to follow below-base forms), */
535 	    if (info[i].indic_position() != POS_BELOW_C &&
536 		(info[i].indic_position() != POS_POST_C || seen_below))
537 	    {
538 	      base = i;
539 	      break;
540 	    }
541 	    if (info[i].indic_position() == POS_BELOW_C)
542 	      seen_below = true;
543 
544 	    /* -> or that is not a pre-base-reordering Ra,
545 	     *
546 	     * IMPLEMENTATION NOTES:
547 	     *
548 	     * Our pre-base-reordering Ra's are marked POS_POST_C, so will be skipped
549 	     * by the logic above already.
550 	     */
551 
552 	    /* -> or arrive at the first consonant. The consonant stopped at will
553 	     * be the base. */
554 	    base = i;
555 	  }
556 	  else
557 	  {
558 	    /* A ZWJ after a Halant stops the base search, and requests an explicit
559 	     * half form.
560 	     * A ZWJ before a Halant, requests a subjoined form instead, and hence
561 	     * search continues.  This is particularly important for Bengali
562 	     * sequence Ra,H,Ya that should form Ya-Phalaa by subjoining Ya. */
563 	    if (start < i &&
564 		info[i].indic_category() == OT_ZWJ &&
565 		info[i - 1].indic_category() == OT_H)
566 	      break;
567 	  }
568 	} while (i > limit);
569       }
570       break;
571 
572       case BASE_POS_LAST_SINHALA:
573       {
574         /* Sinhala base positioning is slightly different from main Indic, in that:
575 	 * 1. Its ZWJ behavior is different,
576 	 * 2. We don't need to look into the font for consonant positions.
577 	 */
578 
579 	if (!has_reph)
580 	  base = limit;
581 
582 	/* Find the last base consonant that is not blocked by ZWJ.  If there is
583 	 * a ZWJ right before a base consonant, that would request a subjoined form. */
584 	for (unsigned int i = limit; i < end; i++)
585 	  if (is_consonant (info[i]))
586 	  {
587 	    if (limit < i && info[i - 1].indic_category() == OT_ZWJ)
588 	      break;
589 	    else
590 	      base = i;
591 	  }
592 
593 	/* Mark all subsequent consonants as below. */
594 	for (unsigned int i = base + 1; i < end; i++)
595 	  if (is_consonant (info[i]))
596 	    info[i].indic_position() = POS_BELOW_C;
597       }
598       break;
599     }
600 
601     /* -> If the syllable starts with Ra + Halant (in a script that has Reph)
602      *    and has more than one consonant, Ra is excluded from candidates for
603      *    base consonants.
604      *
605      *  Only do this for unforced Reph. (ie. not for Ra,H,ZWJ. */
606     if (has_reph && base == start && limit - base <= 2) {
607       /* Have no other consonant, so Reph is not formed and Ra becomes base. */
608       has_reph = false;
609     }
610   }
611 
612 
613   /* 2. Decompose and reorder Matras:
614    *
615    * Each matra and any syllable modifier sign in the syllable are moved to the
616    * appropriate position relative to the consonant(s) in the syllable. The
617    * shaping engine decomposes two- or three-part matras into their constituent
618    * parts before any repositioning. Matra characters are classified by which
619    * consonant in a conjunct they have affinity for and are reordered to the
620    * following positions:
621    *
622    *   o Before first half form in the syllable
623    *   o After subjoined consonants
624    *   o After post-form consonant
625    *   o After main consonant (for above marks)
626    *
627    * IMPLEMENTATION NOTES:
628    *
629    * The normalize() routine has already decomposed matras for us, so we don't
630    * need to worry about that.
631    */
632 
633 
634   /* 3.  Reorder marks to canonical order:
635    *
636    * Adjacent nukta and halant or nukta and vedic sign are always repositioned
637    * if necessary, so that the nukta is first.
638    *
639    * IMPLEMENTATION NOTES:
640    *
641    * We don't need to do this: the normalize() routine already did this for us.
642    */
643 
644 
645   /* Reorder characters */
646 
647   for (unsigned int i = start; i < base; i++)
648     info[i].indic_position() = MIN (POS_PRE_C, (indic_position_t) info[i].indic_position());
649 
650   if (base < end)
651     info[base].indic_position() = POS_BASE_C;
652 
653   /* Mark final consonants.  A final consonant is one appearing after a matra.
654    * Happens in Sinhala. */
655   for (unsigned int i = base + 1; i < end; i++)
656     if (info[i].indic_category() == OT_M) {
657       for (unsigned int j = i + 1; j < end; j++)
658 	if (is_consonant (info[j])) {
659 	 info[j].indic_position() = POS_FINAL_C;
660 	 break;
661        }
662       break;
663     }
664 
665   /* Handle beginning Ra */
666   if (has_reph)
667     info[start].indic_position() = POS_RA_TO_BECOME_REPH;
668 
669   /* For old-style Indic script tags, move the first post-base Halant after
670    * last consonant.
671    *
672    * Reports suggest that in some scripts Uniscribe does this only if there
673    * is *not* a Halant after last consonant already.  We know that is the
674    * case for Kannada, while it reorders unconditionally in other scripts,
675    * eg. Malayalam, Bengali, and Devanagari.  We don't currently know about
676    * other scripts, so we blacklist Kannada.
677    *
678    * Kannada test case:
679    * U+0C9A,U+0CCD,U+0C9A,U+0CCD
680    * With some versions of Lohit Kannada.
681    * https://bugs.freedesktop.org/show_bug.cgi?id=59118
682    *
683    * Malayalam test case:
684    * U+0D38,U+0D4D,U+0D31,U+0D4D,U+0D31,U+0D4D
685    * With lohit-ttf-20121122/Lohit-Malayalam.ttf
686    *
687    * Bengali test case:
688    * U+0998,U+09CD,U+09AF,U+09CD
689    * With Windows XP vrinda.ttf
690    * https://github.com/harfbuzz/harfbuzz/issues/1073
691    *
692    * Devanagari test case:
693    * U+091F,U+094D,U+0930,U+094D
694    * With chandas.ttf
695    * https://github.com/harfbuzz/harfbuzz/issues/1071
696    */
697   if (indic_plan->is_old_spec)
698   {
699     bool disallow_double_halants = buffer->props.script == HB_SCRIPT_KANNADA;
700     for (unsigned int i = base + 1; i < end; i++)
701       if (info[i].indic_category() == OT_H)
702       {
703         unsigned int j;
704         for (j = end - 1; j > i; j--)
705 	  if (is_consonant (info[j]) ||
706 	      (disallow_double_halants && info[j].indic_category() == OT_H))
707 	    break;
708 	if (info[j].indic_category() != OT_H && j > i) {
709 	  /* Move Halant to after last consonant. */
710 	  hb_glyph_info_t t = info[i];
711 	  memmove (&info[i], &info[i + 1], (j - i) * sizeof (info[0]));
712 	  info[j] = t;
713 	}
714         break;
715       }
716   }
717 
718   /* Attach misc marks to previous char to move with them. */
719   {
720     indic_position_t last_pos = POS_START;
721     for (unsigned int i = start; i < end; i++)
722     {
723       if ((FLAG_UNSAFE (info[i].indic_category()) & (JOINER_FLAGS | FLAG (OT_N) | FLAG (OT_RS) | FLAG (OT_H))))
724       {
725 	info[i].indic_position() = last_pos;
726 	if (unlikely (info[i].indic_category() == OT_H &&
727 		      info[i].indic_position() == POS_PRE_M))
728 	{
729 	  /*
730 	   * Uniscribe doesn't move the Halant with Left Matra.
731 	   * TEST: U+092B,U+093F,U+094DE
732 	   * We follow.  This is important for the Sinhala
733 	   * U+0DDA split matra since it decomposes to U+0DD9,U+0DCA
734 	   * where U+0DD9 is a left matra and U+0DCA is the virama.
735 	   * We don't want to move the virama with the left matra.
736 	   * TEST: U+0D9A,U+0DDA
737 	   */
738 	  for (unsigned int j = i; j > start; j--)
739 	    if (info[j - 1].indic_position() != POS_PRE_M) {
740 	      info[i].indic_position() = info[j - 1].indic_position();
741 	      break;
742 	    }
743 	}
744       } else if (info[i].indic_position() != POS_SMVD) {
745         last_pos = (indic_position_t) info[i].indic_position();
746       }
747     }
748   }
749   /* For post-base consonants let them own anything before them
750    * since the last consonant or matra. */
751   {
752     unsigned int last = base;
753     for (unsigned int i = base + 1; i < end; i++)
754       if (is_consonant (info[i]))
755       {
756 	for (unsigned int j = last + 1; j < i; j++)
757 	  if (info[j].indic_position() < POS_SMVD)
758 	    info[j].indic_position() = info[i].indic_position();
759 	last = i;
760       } else if (info[i].indic_category() == OT_M)
761         last = i;
762   }
763 
764 
765   {
766     /* Use syllable() for sort accounting temporarily. */
767     unsigned int syllable = info[start].syllable();
768     for (unsigned int i = start; i < end; i++)
769       info[i].syllable() = i - start;
770 
771     /* Sit tight, rock 'n roll! */
772     hb_stable_sort (info + start, end - start, compare_indic_order);
773     /* Find base again */
774     base = end;
775     for (unsigned int i = start; i < end; i++)
776       if (info[i].indic_position() == POS_BASE_C)
777       {
778 	base = i;
779 	break;
780       }
781     /* Things are out-of-control for post base positions, they may shuffle
782      * around like crazy.  In old-spec mode, we move halants around, so in
783      * that case merge all clusters after base.  Otherwise, check the sort
784      * order and merge as needed.
785      * For pre-base stuff, we handle cluster issues in final reordering.
786      *
787      * We could use buffer->sort() for this, if there was no special
788      * reordering of pre-base stuff happening later...
789      * We don't want to merge_clusters all of that, which buffer->sort()
790      * would.
791      */
792     if (indic_plan->is_old_spec || end - start > 127)
793       buffer->merge_clusters (base, end);
794     else
795     {
796       /* Note!  syllable() is a one-byte field. */
797       for (unsigned int i = base; i < end; i++)
798         if (info[i].syllable() != 255)
799 	{
800 	  unsigned int max = i;
801 	  unsigned int j = start + info[i].syllable();
802 	  while (j != i)
803 	  {
804 	    max = MAX (max, j);
805 	    unsigned int next = start + info[j].syllable();
806 	    info[j].syllable() = 255; /* So we don't process j later again. */
807 	    j = next;
808 	  }
809 	  if (i != max)
810 	    buffer->merge_clusters (i, max + 1);
811 	}
812     }
813 
814     /* Put syllable back in. */
815     for (unsigned int i = start; i < end; i++)
816       info[i].syllable() = syllable;
817   }
818 
819   /* Setup masks now */
820 
821   {
822     hb_mask_t mask;
823 
824     /* Reph */
825     for (unsigned int i = start; i < end && info[i].indic_position() == POS_RA_TO_BECOME_REPH; i++)
826       info[i].mask |= indic_plan->mask_array[RPHF];
827 
828     /* Pre-base */
829     mask = indic_plan->mask_array[HALF];
830     if (!indic_plan->is_old_spec &&
831 	indic_plan->config->blwf_mode == BLWF_MODE_PRE_AND_POST)
832       mask |= indic_plan->mask_array[BLWF];
833     for (unsigned int i = start; i < base; i++)
834       info[i].mask  |= mask;
835     /* Base */
836     mask = 0;
837     if (base < end)
838       info[base].mask |= mask;
839     /* Post-base */
840     mask = indic_plan->mask_array[BLWF] | indic_plan->mask_array[ABVF] | indic_plan->mask_array[PSTF];
841     for (unsigned int i = base + 1; i < end; i++)
842       info[i].mask  |= mask;
843   }
844 
845   if (indic_plan->is_old_spec &&
846       buffer->props.script == HB_SCRIPT_DEVANAGARI)
847   {
848     /* Old-spec eye-lash Ra needs special handling.  From the
849      * spec:
850      *
851      * "The feature 'below-base form' is applied to consonants
852      * having below-base forms and following the base consonant.
853      * The exception is vattu, which may appear below half forms
854      * as well as below the base glyph. The feature 'below-base
855      * form' will be applied to all such occurrences of Ra as well."
856      *
857      * Test case: U+0924,U+094D,U+0930,U+094d,U+0915
858      * with Sanskrit 2003 font.
859      *
860      * However, note that Ra,Halant,ZWJ is the correct way to
861      * request eyelash form of Ra, so we wouldbn't inhibit it
862      * in that sequence.
863      *
864      * Test case: U+0924,U+094D,U+0930,U+094d,U+200D,U+0915
865      */
866     for (unsigned int i = start; i + 1 < base; i++)
867       if (info[i  ].indic_category() == OT_Ra &&
868 	  info[i+1].indic_category() == OT_H  &&
869 	  (i + 2 == base ||
870 	   info[i+2].indic_category() != OT_ZWJ))
871       {
872 	info[i  ].mask |= indic_plan->mask_array[BLWF];
873 	info[i+1].mask |= indic_plan->mask_array[BLWF];
874       }
875   }
876 
877   unsigned int pref_len = 2;
878   if (indic_plan->mask_array[PREF] && base + pref_len < end)
879   {
880     /* Find a Halant,Ra sequence and mark it for pre-base-reordering processing. */
881     for (unsigned int i = base + 1; i + pref_len - 1 < end; i++) {
882       hb_codepoint_t glyphs[2];
883       for (unsigned int j = 0; j < pref_len; j++)
884         glyphs[j] = info[i + j].codepoint;
885       if (indic_plan->pref.would_substitute (glyphs, pref_len, face))
886       {
887 	for (unsigned int j = 0; j < pref_len; j++)
888 	  info[i++].mask |= indic_plan->mask_array[PREF];
889 	break;
890       }
891     }
892   }
893 
894   /* Apply ZWJ/ZWNJ effects */
895   for (unsigned int i = start + 1; i < end; i++)
896     if (is_joiner (info[i])) {
897       bool non_joiner = info[i].indic_category() == OT_ZWNJ;
898       unsigned int j = i;
899 
900       do {
901 	j--;
902 
903 	/* ZWJ/ZWNJ should disable CJCT.  They do that by simply
904 	 * being there, since we don't skip them for the CJCT
905 	 * feature (ie. F_MANUAL_ZWJ) */
906 
907 	/* A ZWNJ disables HALF. */
908 	if (non_joiner)
909 	  info[j].mask &= ~indic_plan->mask_array[HALF];
910 
911       } while (j > start && !is_consonant (info[j]));
912     }
913 }
914 
915 static void
initial_reordering_standalone_cluster(const hb_ot_shape_plan_t * plan,hb_face_t * face,hb_buffer_t * buffer,unsigned int start,unsigned int end)916 initial_reordering_standalone_cluster (const hb_ot_shape_plan_t *plan,
917 				       hb_face_t *face,
918 				       hb_buffer_t *buffer,
919 				       unsigned int start, unsigned int end)
920 {
921   const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) plan->data;
922 
923   /* We treat placeholder/dotted-circle as if they are consonants, so we
924    * should just chain.  Only if not in compatibility mode that is... */
925 
926   if (indic_plan->uniscribe_bug_compatible)
927   {
928     /* For dotted-circle, this is what Uniscribe does:
929      * If dotted-circle is the last glyph, it just does nothing.
930      * Ie. It doesn't form Reph. */
931     if (buffer->info[end - 1].indic_category() == OT_DOTTEDCIRCLE)
932       return;
933   }
934 
935   initial_reordering_consonant_syllable (plan, face, buffer, start, end);
936 }
937 
938 static void
initial_reordering_syllable(const hb_ot_shape_plan_t * plan,hb_face_t * face,hb_buffer_t * buffer,unsigned int start,unsigned int end)939 initial_reordering_syllable (const hb_ot_shape_plan_t *plan,
940 			     hb_face_t *face,
941 			     hb_buffer_t *buffer,
942 			     unsigned int start, unsigned int end)
943 {
944   syllable_type_t syllable_type = (syllable_type_t) (buffer->info[start].syllable() & 0x0F);
945   switch (syllable_type)
946   {
947     case vowel_syllable: /* We made the vowels look like consonants.  So let's call the consonant logic! */
948     case consonant_syllable:
949      initial_reordering_consonant_syllable (plan, face, buffer, start, end);
950      break;
951 
952     case broken_cluster: /* We already inserted dotted-circles, so just call the standalone_cluster. */
953     case standalone_cluster:
954      initial_reordering_standalone_cluster (plan, face, buffer, start, end);
955      break;
956 
957     case symbol_cluster:
958     case non_indic_cluster:
959       break;
960   }
961 }
962 
963 static inline void
insert_dotted_circles(const hb_ot_shape_plan_t * plan HB_UNUSED,hb_font_t * font,hb_buffer_t * buffer)964 insert_dotted_circles (const hb_ot_shape_plan_t *plan HB_UNUSED,
965 		       hb_font_t *font,
966 		       hb_buffer_t *buffer)
967 {
968   /* Note: This loop is extra overhead, but should not be measurable.
969    * TODO Use a buffer scratch flag to remove the loop. */
970   bool has_broken_syllables = false;
971   unsigned int count = buffer->len;
972   hb_glyph_info_t *info = buffer->info;
973   for (unsigned int i = 0; i < count; i++)
974     if ((info[i].syllable() & 0x0F) == broken_cluster)
975     {
976       has_broken_syllables = true;
977       break;
978     }
979   if (likely (!has_broken_syllables))
980     return;
981 
982 
983   hb_codepoint_t dottedcircle_glyph;
984   if (!font->get_nominal_glyph (0x25CCu, &dottedcircle_glyph))
985     return;
986 
987   hb_glyph_info_t dottedcircle = {0};
988   dottedcircle.codepoint = 0x25CCu;
989   set_indic_properties (dottedcircle);
990   dottedcircle.codepoint = dottedcircle_glyph;
991 
992   buffer->clear_output ();
993 
994   buffer->idx = 0;
995   unsigned int last_syllable = 0;
996   while (buffer->idx < buffer->len && buffer->successful)
997   {
998     unsigned int syllable = buffer->cur().syllable();
999     syllable_type_t syllable_type = (syllable_type_t) (syllable & 0x0F);
1000     if (unlikely (last_syllable != syllable && syllable_type == broken_cluster))
1001     {
1002       last_syllable = syllable;
1003 
1004       hb_glyph_info_t ginfo = dottedcircle;
1005       ginfo.cluster = buffer->cur().cluster;
1006       ginfo.mask = buffer->cur().mask;
1007       ginfo.syllable() = buffer->cur().syllable();
1008       /* TODO Set glyph_props? */
1009 
1010       /* Insert dottedcircle after possible Repha. */
1011       while (buffer->idx < buffer->len && buffer->successful &&
1012 	     last_syllable == buffer->cur().syllable() &&
1013 	     buffer->cur().indic_category() == OT_Repha)
1014         buffer->next_glyph ();
1015 
1016       buffer->output_info (ginfo);
1017     }
1018     else
1019       buffer->next_glyph ();
1020   }
1021   buffer->swap_buffers ();
1022 }
1023 
1024 static void
initial_reordering(const hb_ot_shape_plan_t * plan,hb_font_t * font,hb_buffer_t * buffer)1025 initial_reordering (const hb_ot_shape_plan_t *plan,
1026 		    hb_font_t *font,
1027 		    hb_buffer_t *buffer)
1028 {
1029   update_consonant_positions (plan, font, buffer);
1030   insert_dotted_circles (plan, font, buffer);
1031 
1032   foreach_syllable (buffer, start, end)
1033     initial_reordering_syllable (plan, font->face, buffer, start, end);
1034 }
1035 
1036 static void
final_reordering_syllable(const hb_ot_shape_plan_t * plan,hb_buffer_t * buffer,unsigned int start,unsigned int end)1037 final_reordering_syllable (const hb_ot_shape_plan_t *plan,
1038 			   hb_buffer_t *buffer,
1039 			   unsigned int start, unsigned int end)
1040 {
1041   const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) plan->data;
1042   hb_glyph_info_t *info = buffer->info;
1043 
1044 
1045   /* This function relies heavily on halant glyphs.  Lots of ligation
1046    * and possibly multiple substitutions happened prior to this
1047    * phase, and that might have messed up our properties.  Recover
1048    * from a particular case of that where we're fairly sure that a
1049    * class of OT_H is desired but has been lost. */
1050   /* We don't call load_virama_glyph(), since we know it's already
1051    * loaded. */
1052   hb_codepoint_t virama_glyph = indic_plan->virama_glyph.get_relaxed ();
1053   if (virama_glyph)
1054   {
1055     for (unsigned int i = start; i < end; i++)
1056       if (info[i].codepoint == virama_glyph &&
1057 	  _hb_glyph_info_ligated (&info[i]) &&
1058 	  _hb_glyph_info_multiplied (&info[i]))
1059       {
1060         /* This will make sure that this glyph passes is_halant() test. */
1061 	info[i].indic_category() = OT_H;
1062 	_hb_glyph_info_clear_ligated_and_multiplied (&info[i]);
1063       }
1064   }
1065 
1066 
1067   /* 4. Final reordering:
1068    *
1069    * After the localized forms and basic shaping forms GSUB features have been
1070    * applied (see below), the shaping engine performs some final glyph
1071    * reordering before applying all the remaining font features to the entire
1072    * syllable.
1073    */
1074 
1075   bool try_pref = !!indic_plan->mask_array[PREF];
1076 
1077   /* Find base again */
1078   unsigned int base;
1079   for (base = start; base < end; base++)
1080     if (info[base].indic_position() >= POS_BASE_C)
1081     {
1082       if (try_pref && base + 1 < end)
1083       {
1084 	for (unsigned int i = base + 1; i < end; i++)
1085 	  if ((info[i].mask & indic_plan->mask_array[PREF]) != 0)
1086 	  {
1087 	    if (!(_hb_glyph_info_substituted (&info[i]) &&
1088 		  _hb_glyph_info_ligated_and_didnt_multiply (&info[i])))
1089 	    {
1090 	      /* Ok, this was a 'pref' candidate but didn't form any.
1091 	       * Base is around here... */
1092 	      base = i;
1093 	      while (base < end && is_halant (info[base]))
1094 		base++;
1095 	      info[base].indic_position() = POS_BASE_C;
1096 
1097 	      try_pref = false;
1098 	    }
1099 	    break;
1100 	  }
1101       }
1102       /* For Malayalam, skip over unformed below- (but NOT post-) forms. */
1103       if (buffer->props.script == HB_SCRIPT_MALAYALAM)
1104       {
1105 	for (unsigned int i = base + 1; i < end; i++)
1106 	{
1107 	  while (i < end && is_joiner (info[i]))
1108 	    i++;
1109 	  if (i == end || !is_halant (info[i]))
1110 	    break;
1111 	  i++; /* Skip halant. */
1112 	  while (i < end && is_joiner (info[i]))
1113 	    i++;
1114 	  if (i < end && is_consonant (info[i]) && info[i].indic_position() == POS_BELOW_C)
1115 	  {
1116 	    base = i;
1117 	    info[base].indic_position() = POS_BASE_C;
1118 	  }
1119 	}
1120       }
1121 
1122       if (start < base && info[base].indic_position() > POS_BASE_C)
1123         base--;
1124       break;
1125     }
1126   if (base == end && start < base &&
1127       is_one_of (info[base - 1], FLAG (OT_ZWJ)))
1128     base--;
1129   if (base < end)
1130     while (start < base &&
1131 	   is_one_of (info[base], (FLAG (OT_N) | FLAG (OT_H))))
1132       base--;
1133 
1134 
1135   /*   o Reorder matras:
1136    *
1137    *     If a pre-base matra character had been reordered before applying basic
1138    *     features, the glyph can be moved closer to the main consonant based on
1139    *     whether half-forms had been formed. Actual position for the matra is
1140    *     defined as “after last standalone halant glyph, after initial matra
1141    *     position and before the main consonant”. If ZWJ or ZWNJ follow this
1142    *     halant, position is moved after it.
1143    *
1144    * IMPLEMENTATION NOTES:
1145    *
1146    * It looks like the last sentence is wrong.  Testing, with Windows 7 Uniscribe
1147    * and Devanagari shows that the behavior is best described as:
1148    *
1149    * "If ZWJ follows this halant, matra is NOT repositioned after this halant.
1150    *  If ZWNJ follows this halant, position is moved after it."
1151    *
1152    * Test case, with Adobe Devanagari or Nirmala UI:
1153    *
1154    *   U+091F,U+094D,U+200C,U+092F,U+093F
1155    *   (Matra moves to the middle, after ZWNJ.)
1156    *
1157    *   U+091F,U+094D,U+200D,U+092F,U+093F
1158    *   (Matra does NOT move, stays to the left.)
1159    *
1160    * https://github.com/harfbuzz/harfbuzz/issues/1070
1161    */
1162 
1163   if (start + 1 < end && start < base) /* Otherwise there can't be any pre-base matra characters. */
1164   {
1165     /* If we lost track of base, alas, position before last thingy. */
1166     unsigned int new_pos = base == end ? base - 2 : base - 1;
1167 
1168     /* Malayalam / Tamil do not have "half" forms or explicit virama forms.
1169      * The glyphs formed by 'half' are Chillus or ligated explicit viramas.
1170      * We want to position matra after them.
1171      */
1172     if (buffer->props.script != HB_SCRIPT_MALAYALAM && buffer->props.script != HB_SCRIPT_TAMIL)
1173     {
1174     search:
1175       while (new_pos > start &&
1176 	     !(is_one_of (info[new_pos], (FLAG (OT_M) | FLAG (OT_H)))))
1177 	new_pos--;
1178 
1179       /* If we found no Halant we are done.
1180        * Otherwise only proceed if the Halant does
1181        * not belong to the Matra itself! */
1182       if (is_halant (info[new_pos]) &&
1183 	  info[new_pos].indic_position() != POS_PRE_M)
1184       {
1185 #if 0 // See comment above
1186 	/* -> If ZWJ or ZWNJ follow this halant, position is moved after it. */
1187 	if (new_pos + 1 < end && is_joiner (info[new_pos + 1]))
1188 	  new_pos++;
1189 #endif
1190 	if (new_pos + 1 < end)
1191 	{
1192 	  /* -> If ZWJ follows this halant, matra is NOT repositioned after this halant. */
1193 	  if (info[new_pos + 1].indic_category() == OT_ZWJ)
1194 	  {
1195 	    /* Keep searching. */
1196 	    if (new_pos > start)
1197 	    {
1198 	      new_pos--;
1199 	      goto search;
1200 	    }
1201 	  }
1202 	  /* -> If ZWNJ follows this halant, position is moved after it. */
1203 	  if (info[new_pos + 1].indic_category() == OT_ZWNJ)
1204 	    new_pos++;
1205 	}
1206       }
1207       else
1208         new_pos = start; /* No move. */
1209     }
1210 
1211     if (start < new_pos && info[new_pos].indic_position () != POS_PRE_M)
1212     {
1213       /* Now go see if there's actually any matras... */
1214       for (unsigned int i = new_pos; i > start; i--)
1215 	if (info[i - 1].indic_position () == POS_PRE_M)
1216 	{
1217 	  unsigned int old_pos = i - 1;
1218 	  if (old_pos < base && base <= new_pos) /* Shouldn't actually happen. */
1219 	    base--;
1220 
1221 	  hb_glyph_info_t tmp = info[old_pos];
1222 	  memmove (&info[old_pos], &info[old_pos + 1], (new_pos - old_pos) * sizeof (info[0]));
1223 	  info[new_pos] = tmp;
1224 
1225 	  /* Note: this merge_clusters() is intentionally *after* the reordering.
1226 	   * Indic matra reordering is special and tricky... */
1227 	  buffer->merge_clusters (new_pos, MIN (end, base + 1));
1228 
1229 	  new_pos--;
1230 	}
1231     } else {
1232       for (unsigned int i = start; i < base; i++)
1233 	if (info[i].indic_position () == POS_PRE_M) {
1234 	  buffer->merge_clusters (i, MIN (end, base + 1));
1235 	  break;
1236 	}
1237     }
1238   }
1239 
1240 
1241   /*   o Reorder reph:
1242    *
1243    *     Reph’s original position is always at the beginning of the syllable,
1244    *     (i.e. it is not reordered at the character reordering stage). However,
1245    *     it will be reordered according to the basic-forms shaping results.
1246    *     Possible positions for reph, depending on the script, are; after main,
1247    *     before post-base consonant forms, and after post-base consonant forms.
1248    */
1249 
1250   /* Two cases:
1251    *
1252    * - If repha is encoded as a sequence of characters (Ra,H or Ra,H,ZWJ), then
1253    *   we should only move it if the sequence ligated to the repha form.
1254    *
1255    * - If repha is encoded separately and in the logical position, we should only
1256    *   move it if it did NOT ligate.  If it ligated, it's probably the font trying
1257    *   to make it work without the reordering.
1258    */
1259   if (start + 1 < end &&
1260       info[start].indic_position() == POS_RA_TO_BECOME_REPH &&
1261       ((info[start].indic_category() == OT_Repha) ^
1262        _hb_glyph_info_ligated_and_didnt_multiply (&info[start])))
1263   {
1264     unsigned int new_reph_pos;
1265     reph_position_t reph_pos = indic_plan->config->reph_pos;
1266 
1267     /*       1. If reph should be positioned after post-base consonant forms,
1268      *          proceed to step 5.
1269      */
1270     if (reph_pos == REPH_POS_AFTER_POST)
1271     {
1272       goto reph_step_5;
1273     }
1274 
1275     /*       2. If the reph repositioning class is not after post-base: target
1276      *          position is after the first explicit halant glyph between the
1277      *          first post-reph consonant and last main consonant. If ZWJ or ZWNJ
1278      *          are following this halant, position is moved after it. If such
1279      *          position is found, this is the target position. Otherwise,
1280      *          proceed to the next step.
1281      *
1282      *          Note: in old-implementation fonts, where classifications were
1283      *          fixed in shaping engine, there was no case where reph position
1284      *          will be found on this step.
1285      */
1286     {
1287       new_reph_pos = start + 1;
1288       while (new_reph_pos < base && !is_halant (info[new_reph_pos]))
1289 	new_reph_pos++;
1290 
1291       if (new_reph_pos < base && is_halant (info[new_reph_pos]))
1292       {
1293 	/* ->If ZWJ or ZWNJ are following this halant, position is moved after it. */
1294 	if (new_reph_pos + 1 < base && is_joiner (info[new_reph_pos + 1]))
1295 	  new_reph_pos++;
1296 	goto reph_move;
1297       }
1298     }
1299 
1300     /*       3. If reph should be repositioned after the main consonant: find the
1301      *          first consonant not ligated with main, or find the first
1302      *          consonant that is not a potential pre-base-reordering Ra.
1303      */
1304     if (reph_pos == REPH_POS_AFTER_MAIN)
1305     {
1306       new_reph_pos = base;
1307       while (new_reph_pos + 1 < end && info[new_reph_pos + 1].indic_position() <= POS_AFTER_MAIN)
1308 	new_reph_pos++;
1309       if (new_reph_pos < end)
1310         goto reph_move;
1311     }
1312 
1313     /*       4. If reph should be positioned before post-base consonant, find
1314      *          first post-base classified consonant not ligated with main. If no
1315      *          consonant is found, the target position should be before the
1316      *          first matra, syllable modifier sign or vedic sign.
1317      */
1318     /* This is our take on what step 4 is trying to say (and failing, BADLY). */
1319     if (reph_pos == REPH_POS_AFTER_SUB)
1320     {
1321       new_reph_pos = base;
1322       while (new_reph_pos + 1 < end &&
1323 	     !( FLAG_UNSAFE (info[new_reph_pos + 1].indic_position()) & (FLAG (POS_POST_C) | FLAG (POS_AFTER_POST) | FLAG (POS_SMVD))))
1324 	new_reph_pos++;
1325       if (new_reph_pos < end)
1326         goto reph_move;
1327     }
1328 
1329     /*       5. If no consonant is found in steps 3 or 4, move reph to a position
1330      *          immediately before the first post-base matra, syllable modifier
1331      *          sign or vedic sign that has a reordering class after the intended
1332      *          reph position. For example, if the reordering position for reph
1333      *          is post-main, it will skip above-base matras that also have a
1334      *          post-main position.
1335      */
1336     reph_step_5:
1337     {
1338       /* Copied from step 2. */
1339       new_reph_pos = start + 1;
1340       while (new_reph_pos < base && !is_halant (info[new_reph_pos]))
1341 	new_reph_pos++;
1342 
1343       if (new_reph_pos < base && is_halant (info[new_reph_pos]))
1344       {
1345 	/* ->If ZWJ or ZWNJ are following this halant, position is moved after it. */
1346 	if (new_reph_pos + 1 < base && is_joiner (info[new_reph_pos + 1]))
1347 	  new_reph_pos++;
1348 	goto reph_move;
1349       }
1350     }
1351 
1352     /*       6. Otherwise, reorder reph to the end of the syllable.
1353      */
1354     {
1355       new_reph_pos = end - 1;
1356       while (new_reph_pos > start && info[new_reph_pos].indic_position() == POS_SMVD)
1357 	new_reph_pos--;
1358 
1359       /*
1360        * If the Reph is to be ending up after a Matra,Halant sequence,
1361        * position it before that Halant so it can interact with the Matra.
1362        * However, if it's a plain Consonant,Halant we shouldn't do that.
1363        * Uniscribe doesn't do this.
1364        * TEST: U+0930,U+094D,U+0915,U+094B,U+094D
1365        */
1366       if (!indic_plan->uniscribe_bug_compatible &&
1367 	  unlikely (is_halant (info[new_reph_pos]))) {
1368 	for (unsigned int i = base + 1; i < new_reph_pos; i++)
1369 	  if (info[i].indic_category() == OT_M) {
1370 	    /* Ok, got it. */
1371 	    new_reph_pos--;
1372 	  }
1373       }
1374       goto reph_move;
1375     }
1376 
1377     reph_move:
1378     {
1379       /* Move */
1380       buffer->merge_clusters (start, new_reph_pos + 1);
1381       hb_glyph_info_t reph = info[start];
1382       memmove (&info[start], &info[start + 1], (new_reph_pos - start) * sizeof (info[0]));
1383       info[new_reph_pos] = reph;
1384 
1385       if (start < base && base <= new_reph_pos)
1386 	base--;
1387     }
1388   }
1389 
1390 
1391   /*   o Reorder pre-base-reordering consonants:
1392    *
1393    *     If a pre-base-reordering consonant is found, reorder it according to
1394    *     the following rules:
1395    */
1396 
1397   if (try_pref && base + 1 < end) /* Otherwise there can't be any pre-base-reordering Ra. */
1398   {
1399     for (unsigned int i = base + 1; i < end; i++)
1400       if ((info[i].mask & indic_plan->mask_array[PREF]) != 0)
1401       {
1402 	/*       1. Only reorder a glyph produced by substitution during application
1403 	 *          of the <pref> feature. (Note that a font may shape a Ra consonant with
1404 	 *          the feature generally but block it in certain contexts.)
1405 	 */
1406         /* Note: We just check that something got substituted.  We don't check that
1407 	 * the <pref> feature actually did it...
1408 	 *
1409 	 * Reorder pref only if it ligated. */
1410 	if (_hb_glyph_info_ligated_and_didnt_multiply (&info[i]))
1411 	{
1412 	  /*
1413 	   *       2. Try to find a target position the same way as for pre-base matra.
1414 	   *          If it is found, reorder pre-base consonant glyph.
1415 	   *
1416 	   *       3. If position is not found, reorder immediately before main
1417 	   *          consonant.
1418 	   */
1419 
1420 	  unsigned int new_pos = base;
1421 	  /* Malayalam / Tamil do not have "half" forms or explicit virama forms.
1422 	   * The glyphs formed by 'half' are Chillus or ligated explicit viramas.
1423 	   * We want to position matra after them.
1424 	   */
1425 	  if (buffer->props.script != HB_SCRIPT_MALAYALAM && buffer->props.script != HB_SCRIPT_TAMIL)
1426 	  {
1427 	    while (new_pos > start &&
1428 		   !(is_one_of (info[new_pos - 1], FLAG(OT_M) | FLAG (OT_H))))
1429 	      new_pos--;
1430 	  }
1431 
1432 	  if (new_pos > start && is_halant (info[new_pos - 1]))
1433 	  {
1434 	    /* -> If ZWJ or ZWNJ follow this halant, position is moved after it. */
1435 	    if (new_pos < end && is_joiner (info[new_pos]))
1436 	      new_pos++;
1437 	  }
1438 
1439 	  {
1440 	    unsigned int old_pos = i;
1441 
1442 	    buffer->merge_clusters (new_pos, old_pos + 1);
1443 	    hb_glyph_info_t tmp = info[old_pos];
1444 	    memmove (&info[new_pos + 1], &info[new_pos], (old_pos - new_pos) * sizeof (info[0]));
1445 	    info[new_pos] = tmp;
1446 
1447 	    if (new_pos <= base && base < old_pos)
1448 	      base++;
1449 	  }
1450 	}
1451 
1452         break;
1453       }
1454   }
1455 
1456 
1457   /* Apply 'init' to the Left Matra if it's a word start. */
1458   if (info[start].indic_position () == POS_PRE_M)
1459   {
1460     if (!start ||
1461 	!(FLAG_UNSAFE (_hb_glyph_info_get_general_category (&info[start - 1])) &
1462 	 FLAG_RANGE (HB_UNICODE_GENERAL_CATEGORY_FORMAT, HB_UNICODE_GENERAL_CATEGORY_NON_SPACING_MARK)))
1463       info[start].mask |= indic_plan->mask_array[INIT];
1464     else
1465       buffer->unsafe_to_break (start - 1, start + 1);
1466   }
1467 
1468 
1469   /*
1470    * Finish off the clusters and go home!
1471    */
1472   if (indic_plan->uniscribe_bug_compatible)
1473   {
1474     switch ((hb_tag_t) plan->props.script)
1475     {
1476       case HB_SCRIPT_TAMIL:
1477       case HB_SCRIPT_SINHALA:
1478         break;
1479 
1480       default:
1481 	/* Uniscribe merges the entire syllable into a single cluster... Except for Tamil & Sinhala.
1482 	 * This means, half forms are submerged into the main consonant's cluster.
1483 	 * This is unnecessary, and makes cursor positioning harder, but that's what
1484 	 * Uniscribe does. */
1485 	buffer->merge_clusters (start, end);
1486 	break;
1487     }
1488   }
1489 }
1490 
1491 
1492 static void
final_reordering(const hb_ot_shape_plan_t * plan,hb_font_t * font HB_UNUSED,hb_buffer_t * buffer)1493 final_reordering (const hb_ot_shape_plan_t *plan,
1494 		  hb_font_t *font HB_UNUSED,
1495 		  hb_buffer_t *buffer)
1496 {
1497   unsigned int count = buffer->len;
1498   if (unlikely (!count)) return;
1499 
1500   foreach_syllable (buffer, start, end)
1501     final_reordering_syllable (plan, buffer, start, end);
1502 
1503   HB_BUFFER_DEALLOCATE_VAR (buffer, indic_category);
1504   HB_BUFFER_DEALLOCATE_VAR (buffer, indic_position);
1505 }
1506 
1507 
1508 static void
clear_syllables(const hb_ot_shape_plan_t * plan HB_UNUSED,hb_font_t * font HB_UNUSED,hb_buffer_t * buffer)1509 clear_syllables (const hb_ot_shape_plan_t *plan HB_UNUSED,
1510 		 hb_font_t *font HB_UNUSED,
1511 		 hb_buffer_t *buffer)
1512 {
1513   hb_glyph_info_t *info = buffer->info;
1514   unsigned int count = buffer->len;
1515   for (unsigned int i = 0; i < count; i++)
1516     info[i].syllable() = 0;
1517 }
1518 
1519 
1520 static void
preprocess_text_indic(const hb_ot_shape_plan_t * plan,hb_buffer_t * buffer,hb_font_t * font)1521 preprocess_text_indic (const hb_ot_shape_plan_t *plan,
1522 		       hb_buffer_t              *buffer,
1523 		       hb_font_t                *font)
1524 {
1525   _hb_preprocess_text_vowel_constraints (plan, buffer, font);
1526 }
1527 
1528 static bool
decompose_indic(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t ab,hb_codepoint_t * a,hb_codepoint_t * b)1529 decompose_indic (const hb_ot_shape_normalize_context_t *c,
1530 		 hb_codepoint_t  ab,
1531 		 hb_codepoint_t *a,
1532 		 hb_codepoint_t *b)
1533 {
1534   switch (ab)
1535   {
1536     /* Don't decompose these. */
1537     case 0x0931u  : return false; /* DEVANAGARI LETTER RRA */
1538     // https://github.com/harfbuzz/harfbuzz/issues/779
1539     case 0x09DCu  : return false; /* BENGALI LETTER RRA */
1540     case 0x09DDu  : return false; /* BENGALI LETTER RHA */
1541     case 0x0B94u  : return false; /* TAMIL LETTER AU */
1542 
1543 
1544     /*
1545      * Decompose split matras that don't have Unicode decompositions.
1546      */
1547 
1548 #if 0
1549     /* Gujarati */
1550     /* This one has no decomposition in Unicode, but needs no decomposition either. */
1551     /* case 0x0AC9u  : return false; */
1552 
1553     /* Oriya */
1554     case 0x0B57u  : *a = no decomp, -> RIGHT; return true;
1555 #endif
1556   }
1557 
1558   if ((ab == 0x0DDAu || hb_in_range<hb_codepoint_t> (ab, 0x0DDCu, 0x0DDEu)))
1559   {
1560     /*
1561      * Sinhala split matras...  Let the fun begin.
1562      *
1563      * These four characters have Unicode decompositions.  However, Uniscribe
1564      * decomposes them "Khmer-style", that is, it uses the character itself to
1565      * get the second half.  The first half of all four decompositions is always
1566      * U+0DD9.
1567      *
1568      * Now, there are buggy fonts, namely, the widely used lklug.ttf, that are
1569      * broken with Uniscribe.  But we need to support them.  As such, we only
1570      * do the Uniscribe-style decomposition if the character is transformed into
1571      * its "sec.half" form by the 'pstf' feature.  Otherwise, we fall back to
1572      * Unicode decomposition.
1573      *
1574      * Note that we can't unconditionally use Unicode decomposition.  That would
1575      * break some other fonts, that are designed to work with Uniscribe, and
1576      * don't have positioning features for the Unicode-style decomposition.
1577      *
1578      * Argh...
1579      *
1580      * The Uniscribe behavior is now documented in the newly published Sinhala
1581      * spec in 2012:
1582      *
1583      *   https://docs.microsoft.com/en-us/typography/script-development/sinhala#shaping
1584      */
1585 
1586     const indic_shape_plan_t *indic_plan = (const indic_shape_plan_t *) c->plan->data;
1587 
1588     hb_codepoint_t glyph;
1589 
1590     if (hb_options ().uniscribe_bug_compatible ||
1591 	(c->font->get_nominal_glyph (ab, &glyph) &&
1592 	 indic_plan->pstf.would_substitute (&glyph, 1, c->font->face)))
1593     {
1594       /* Ok, safe to use Uniscribe-style decomposition. */
1595       *a = 0x0DD9u;
1596       *b = ab;
1597       return true;
1598     }
1599   }
1600 
1601   return (bool) c->unicode->decompose (ab, a, b);
1602 }
1603 
1604 static bool
compose_indic(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t a,hb_codepoint_t b,hb_codepoint_t * ab)1605 compose_indic (const hb_ot_shape_normalize_context_t *c,
1606 	       hb_codepoint_t  a,
1607 	       hb_codepoint_t  b,
1608 	       hb_codepoint_t *ab)
1609 {
1610   /* Avoid recomposing split matras. */
1611   if (HB_UNICODE_GENERAL_CATEGORY_IS_MARK (c->unicode->general_category (a)))
1612     return false;
1613 
1614   /* Composition-exclusion exceptions that we want to recompose. */
1615   if (a == 0x09AFu && b == 0x09BCu) { *ab = 0x09DFu; return true; }
1616 
1617   return (bool) c->unicode->compose (a, b, ab);
1618 }
1619 
1620 
1621 const hb_ot_complex_shaper_t _hb_ot_complex_shaper_indic =
1622 {
1623   collect_features_indic,
1624   override_features_indic,
1625   data_create_indic,
1626   data_destroy_indic,
1627   preprocess_text_indic,
1628   nullptr, /* postprocess_glyphs */
1629   HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT,
1630   decompose_indic,
1631   compose_indic,
1632   setup_masks_indic,
1633   HB_TAG_NONE, /* gpos_tag */
1634   nullptr, /* reorder_marks */
1635   HB_OT_SHAPE_ZERO_WIDTH_MARKS_NONE,
1636   false, /* fallback_position */
1637 };
1638