1 /*
2  * Copyright © 2011,2012  Google, Inc.
3  *
4  *  This is part of HarfBuzz, a text shaping library.
5  *
6  * Permission is hereby granted, without written agreement and without
7  * license or royalty fees, to use, copy, modify, and distribute this
8  * software and its documentation for any purpose, provided that the
9  * above copyright notice and the following two paragraphs appear in
10  * all copies of this software.
11  *
12  * IN NO EVENT SHALL THE COPYRIGHT HOLDER BE LIABLE TO ANY PARTY FOR
13  * DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL DAMAGES
14  * ARISING OUT OF THE USE OF THIS SOFTWARE AND ITS DOCUMENTATION, EVEN
15  * IF THE COPYRIGHT HOLDER HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH
16  * DAMAGE.
17  *
18  * THE COPYRIGHT HOLDER SPECIFICALLY DISCLAIMS ANY WARRANTIES, INCLUDING,
19  * BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
20  * FITNESS FOR A PARTICULAR PURPOSE.  THE SOFTWARE PROVIDED HEREUNDER IS
21  * ON AN "AS IS" BASIS, AND THE COPYRIGHT HOLDER HAS NO OBLIGATION TO
22  * PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.
23  *
24  * Google Author(s): Behdad Esfahbod
25  */
26 
27 #include "hb.hh"
28 
29 #ifndef HB_NO_OT_SHAPE
30 
31 #include "hb-ot-shape-normalize.hh"
32 #include "hb-ot-shape-complex.hh"
33 #include "hb-ot-shape.hh"
34 
35 
36 /*
37  * HIGHLEVEL DESIGN:
38  *
39  * This file exports one main function: _hb_ot_shape_normalize().
40  *
41  * This function closely reflects the Unicode Normalization Algorithm,
42  * yet it's different.
43  *
44  * Each shaper specifies whether it prefers decomposed (NFD) or composed (NFC).
45  * The logic however tries to use whatever the font can support.
46  *
47  * In general what happens is that: each grapheme is decomposed in a chain
48  * of 1:2 decompositions, marks reordered, and then recomposed if desired,
49  * so far it's like Unicode Normalization.  However, the decomposition and
50  * recomposition only happens if the font supports the resulting characters.
51  *
52  * The goals are:
53  *
54  *   - Try to render all canonically equivalent strings similarly.  To really
55  *     achieve this we have to always do the full decomposition and then
56  *     selectively recompose from there.  It's kinda too expensive though, so
57  *     we skip some cases.  For example, if composed is desired, we simply
58  *     don't touch 1-character clusters that are supported by the font, even
59  *     though their NFC may be different.
60  *
61  *   - When a font has a precomposed character for a sequence but the 'ccmp'
62  *     feature in the font is not adequate, use the precomposed character
63  *     which typically has better mark positioning.
64  *
65  *   - When a font does not support a combining mark, but supports it precomposed
66  *     with previous base, use that.  This needs the itemizer to have this
67  *     knowledge too.  We need to provide assistance to the itemizer.
68  *
69  *   - When a font does not support a character but supports its canonical
70  *     decomposition, well, use the decomposition.
71  *
72  *   - The complex shapers can customize the compose and decompose functions to
73  *     offload some of their requirements to the normalizer.  For example, the
74  *     Indic shaper may want to disallow recomposing of two matras.
75  */
76 
77 static bool
decompose_unicode(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t ab,hb_codepoint_t * a,hb_codepoint_t * b)78 decompose_unicode (const hb_ot_shape_normalize_context_t *c,
79 		   hb_codepoint_t  ab,
80 		   hb_codepoint_t *a,
81 		   hb_codepoint_t *b)
82 {
83   return (bool) c->unicode->decompose (ab, a, b);
84 }
85 
86 static bool
compose_unicode(const hb_ot_shape_normalize_context_t * c,hb_codepoint_t a,hb_codepoint_t b,hb_codepoint_t * ab)87 compose_unicode (const hb_ot_shape_normalize_context_t *c,
88 		 hb_codepoint_t  a,
89 		 hb_codepoint_t  b,
90 		 hb_codepoint_t *ab)
91 {
92   return (bool) c->unicode->compose (a, b, ab);
93 }
94 
95 static inline void
set_glyph(hb_glyph_info_t & info,hb_font_t * font)96 set_glyph (hb_glyph_info_t &info, hb_font_t *font)
97 {
98   (void) font->get_nominal_glyph (info.codepoint, &info.glyph_index());
99 }
100 
101 static inline void
output_char(hb_buffer_t * buffer,hb_codepoint_t unichar,hb_codepoint_t glyph)102 output_char (hb_buffer_t *buffer, hb_codepoint_t unichar, hb_codepoint_t glyph)
103 {
104   buffer->cur().glyph_index() = glyph;
105   buffer->output_glyph (unichar); /* This is very confusing indeed. */
106   _hb_glyph_info_set_unicode_props (&buffer->prev(), buffer);
107 }
108 
109 static inline void
next_char(hb_buffer_t * buffer,hb_codepoint_t glyph)110 next_char (hb_buffer_t *buffer, hb_codepoint_t glyph)
111 {
112   buffer->cur().glyph_index() = glyph;
113   buffer->next_glyph ();
114 }
115 
116 static inline void
skip_char(hb_buffer_t * buffer)117 skip_char (hb_buffer_t *buffer)
118 {
119   buffer->skip_glyph ();
120 }
121 
122 /* Returns 0 if didn't decompose, number of resulting characters otherwise. */
123 static inline unsigned int
decompose(const hb_ot_shape_normalize_context_t * c,bool shortest,hb_codepoint_t ab)124 decompose (const hb_ot_shape_normalize_context_t *c, bool shortest, hb_codepoint_t ab)
125 {
126   hb_codepoint_t a = 0, b = 0, a_glyph = 0, b_glyph = 0;
127   hb_buffer_t * const buffer = c->buffer;
128   hb_font_t * const font = c->font;
129 
130   if (!c->decompose (c, ab, &a, &b) ||
131       (b && !font->get_nominal_glyph (b, &b_glyph)))
132     return 0;
133 
134   bool has_a = (bool) font->get_nominal_glyph (a, &a_glyph);
135   if (shortest && has_a) {
136     /* Output a and b */
137     output_char (buffer, a, a_glyph);
138     if (likely (b)) {
139       output_char (buffer, b, b_glyph);
140       return 2;
141     }
142     return 1;
143   }
144 
145   unsigned int ret;
146   if ((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 
166 static inline void
decompose_current_character(const hb_ot_shape_normalize_context_t * c,bool shortest)167 decompose_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))
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))
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 && c->font->get_nominal_glyph (0x0020u, &space_glyph))
196     {
197       _hb_glyph_info_set_unicode_space_fallback_type (&buffer->cur(), space_type);
198       next_char (buffer, space_glyph);
199       buffer->scratch_flags |= HB_BUFFER_SCRATCH_FLAG_HAS_SPACE_FALLBACK;
200       return;
201     }
202   }
203 
204   if (u == 0x2011u)
205   {
206     /* U+2011 is the only sensible character that is a no-break version of another character
207      * and not a space.  The space ones are handled already.  Handle this lone one. */
208     hb_codepoint_t other_glyph;
209     if (c->font->get_nominal_glyph (0x2010u, &other_glyph))
210     {
211       next_char (buffer, other_glyph);
212       return;
213     }
214   }
215 
216   next_char (buffer, glyph); /* glyph is initialized in earlier branches. */
217 }
218 
219 static inline void
handle_variation_selector_cluster(const hb_ot_shape_normalize_context_t * c,unsigned int end,bool short_circuit HB_UNUSED)220 handle_variation_selector_cluster (const hb_ot_shape_normalize_context_t *c,
221 				   unsigned int end,
222 				   bool short_circuit HB_UNUSED)
223 {
224   /* TODO Currently if there's a variation-selector we give-up, it's just too hard. */
225   hb_buffer_t * const buffer = c->buffer;
226   hb_font_t * const font = c->font;
227   for (; buffer->idx < end - 1 && buffer->successful;) {
228     if (unlikely (buffer->unicode->is_variation_selector (buffer->cur(+1).codepoint))) {
229       if (font->get_variation_glyph (buffer->cur().codepoint, buffer->cur(+1).codepoint, &buffer->cur().glyph_index()))
230       {
231 	hb_codepoint_t unicode = buffer->cur().codepoint;
232 	buffer->replace_glyphs (2, 1, &unicode);
233       }
234       else
235       {
236 	/* Just pass on the two characters separately, let GSUB do its magic. */
237 	set_glyph (buffer->cur(), font);
238 	buffer->next_glyph ();
239 	set_glyph (buffer->cur(), font);
240 	buffer->next_glyph ();
241       }
242       /* Skip any further variation selectors. */
243       while (buffer->idx < end && unlikely (buffer->unicode->is_variation_selector (buffer->cur().codepoint)))
244       {
245 	set_glyph (buffer->cur(), font);
246 	buffer->next_glyph ();
247       }
248     } else {
249       set_glyph (buffer->cur(), font);
250       buffer->next_glyph ();
251     }
252   }
253   if (likely (buffer->idx < end)) {
254     set_glyph (buffer->cur(), font);
255     buffer->next_glyph ();
256   }
257 }
258 
259 static inline void
decompose_multi_char_cluster(const hb_ot_shape_normalize_context_t * c,unsigned int end,bool short_circuit)260 decompose_multi_char_cluster (const hb_ot_shape_normalize_context_t *c, unsigned int end, bool short_circuit)
261 {
262   hb_buffer_t * const buffer = c->buffer;
263   for (unsigned int i = buffer->idx; i < end && buffer->successful; i++)
264     if (unlikely (buffer->unicode->is_variation_selector (buffer->info[i].codepoint))) {
265       handle_variation_selector_cluster (c, end, short_circuit);
266       return;
267     }
268 
269   while (buffer->idx < end && buffer->successful)
270     decompose_current_character (c, short_circuit);
271 }
272 
273 
274 static int
compare_combining_class(const hb_glyph_info_t * pa,const hb_glyph_info_t * pb)275 compare_combining_class (const hb_glyph_info_t *pa, const hb_glyph_info_t *pb)
276 {
277   unsigned int a = _hb_glyph_info_get_modified_combining_class (pa);
278   unsigned int b = _hb_glyph_info_get_modified_combining_class (pb);
279 
280   return a < b ? -1 : a == b ? 0 : +1;
281 }
282 
283 
284 void
_hb_ot_shape_normalize(const hb_ot_shape_plan_t * plan,hb_buffer_t * buffer,hb_font_t * font)285 _hb_ot_shape_normalize (const hb_ot_shape_plan_t *plan,
286 			hb_buffer_t *buffer,
287 			hb_font_t *font)
288 {
289   if (unlikely (!buffer->len)) return;
290 
291   _hb_buffer_assert_unicode_vars (buffer);
292 
293   hb_ot_shape_normalization_mode_t mode = plan->shaper->normalization_preference;
294   if (mode == HB_OT_SHAPE_NORMALIZATION_MODE_AUTO)
295   {
296     if (plan->has_gpos_mark)
297       // https://github.com/harfbuzz/harfbuzz/issues/653#issuecomment-423905920
298       //mode = HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED;
299       mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
300     else
301       mode = HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS;
302   }
303 
304   const hb_ot_shape_normalize_context_t c = {
305     plan,
306     buffer,
307     font,
308     buffer->unicode,
309     plan->shaper->decompose ? plan->shaper->decompose : decompose_unicode,
310     plan->shaper->compose   ? plan->shaper->compose   : compose_unicode
311   };
312 
313   bool always_short_circuit = mode == HB_OT_SHAPE_NORMALIZATION_MODE_NONE;
314   bool might_short_circuit = always_short_circuit ||
315 			     (mode != HB_OT_SHAPE_NORMALIZATION_MODE_DECOMPOSED &&
316 			      mode != HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT);
317   unsigned int count;
318 
319   /* We do a fairly straightforward yet custom normalization process in three
320    * separate rounds: decompose, reorder, recompose (if desired).  Currently
321    * this makes two buffer swaps.  We can make it faster by moving the last
322    * two rounds into the inner loop for the first round, but it's more readable
323    * this way. */
324 
325 
326   /* First round, decompose */
327 
328   bool all_simple = true;
329   {
330     buffer->clear_output ();
331     count = buffer->len;
332     buffer->idx = 0;
333     do
334     {
335       unsigned int end;
336       for (end = buffer->idx + 1; end < count; end++)
337 	if (unlikely (HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->info[end]))))
338 	  break;
339 
340       if (end < count)
341 	end--; /* Leave one base for the marks to cluster with. */
342 
343       /* From idx to end are simple clusters. */
344       if (might_short_circuit)
345       {
346 	unsigned int done = font->get_nominal_glyphs (end - buffer->idx,
347 						      &buffer->cur().codepoint,
348 						      sizeof (buffer->info[0]),
349 						      &buffer->cur().glyph_index(),
350 						      sizeof (buffer->info[0]));
351 	buffer->next_glyphs (done);
352       }
353       while (buffer->idx < end && buffer->successful)
354 	decompose_current_character (&c, might_short_circuit);
355 
356       if (buffer->idx == count || !buffer->successful)
357 	break;
358 
359       all_simple = false;
360 
361       /* Find all the marks now. */
362       for (end = buffer->idx + 1; end < count; end++)
363 	if (!HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->info[end])))
364 	  break;
365 
366       /* idx to end is one non-simple cluster. */
367       decompose_multi_char_cluster (&c, end, always_short_circuit);
368     }
369     while (buffer->idx < count && buffer->successful);
370     buffer->swap_buffers ();
371   }
372 
373 
374   /* Second round, reorder (inplace) */
375 
376   if (!all_simple)
377   {
378     count = buffer->len;
379     for (unsigned int i = 0; i < count; i++)
380     {
381       if (_hb_glyph_info_get_modified_combining_class (&buffer->info[i]) == 0)
382 	continue;
383 
384       unsigned int end;
385       for (end = i + 1; end < count; end++)
386 	if (_hb_glyph_info_get_modified_combining_class (&buffer->info[end]) == 0)
387 	  break;
388 
389       /* We are going to do a O(n^2).  Only do this if the sequence is short. */
390       if (end - i > HB_OT_SHAPE_COMPLEX_MAX_COMBINING_MARKS) {
391 	i = end;
392 	continue;
393       }
394 
395       buffer->sort (i, end, compare_combining_class);
396 
397       if (plan->shaper->reorder_marks)
398 	plan->shaper->reorder_marks (plan, buffer, i, end);
399 
400       i = end;
401     }
402   }
403   if (buffer->scratch_flags & HB_BUFFER_SCRATCH_FLAG_HAS_CGJ)
404   {
405     /* For all CGJ, check if it prevented any reordering at all.
406      * If it did NOT, then make it skippable.
407      * https://github.com/harfbuzz/harfbuzz/issues/554
408      */
409     for (unsigned int i = 1; i + 1 < buffer->len; i++)
410       if (buffer->info[i].codepoint == 0x034Fu/*CGJ*/ &&
411 	  info_cc(buffer->info[i-1]) <= info_cc(buffer->info[i+1]))
412       {
413 	_hb_glyph_info_unhide (&buffer->info[i]);
414       }
415   }
416 
417 
418   /* Third round, recompose */
419 
420   if (!all_simple &&
421       (mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS ||
422        mode == HB_OT_SHAPE_NORMALIZATION_MODE_COMPOSED_DIACRITICS_NO_SHORT_CIRCUIT))
423   {
424     /* As noted in the comment earlier, we don't try to combine
425      * ccc=0 chars with their previous Starter. */
426 
427     buffer->clear_output ();
428     count = buffer->len;
429     unsigned int starter = 0;
430     buffer->next_glyph ();
431     while (buffer->idx < count && buffer->successful)
432     {
433       hb_codepoint_t composed, glyph;
434       if (/* We don't try to compose a non-mark character with it's preceding starter.
435 	   * This is both an optimization to avoid trying to compose every two neighboring
436 	   * glyphs in most scripts AND a desired feature for Hangul.  Apparently Hangul
437 	   * fonts are not designed to mix-and-match pre-composed syllables and Jamo. */
438 	  HB_UNICODE_GENERAL_CATEGORY_IS_MARK (_hb_glyph_info_get_general_category (&buffer->cur())))
439       {
440 	if (/* If there's anything between the starter and this char, they should have CCC
441 	     * smaller than this character's. */
442 	    (starter == buffer->out_len - 1 ||
443 	     info_cc (buffer->prev()) < info_cc (buffer->cur())) &&
444 	    /* And compose. */
445 	    c.compose (&c,
446 		       buffer->out_info[starter].codepoint,
447 		       buffer->cur().codepoint,
448 		       &composed) &&
449 	    /* And the font has glyph for the composite. */
450 	    font->get_nominal_glyph (composed, &glyph))
451 	{
452 	  /* Composes. */
453 	  buffer->next_glyph (); /* Copy to out-buffer. */
454 	  if (unlikely (!buffer->successful))
455 	    return;
456 	  buffer->merge_out_clusters (starter, buffer->out_len);
457 	  buffer->out_len--; /* Remove the second composable. */
458 	  /* Modify starter and carry on. */
459 	  buffer->out_info[starter].codepoint = composed;
460 	  buffer->out_info[starter].glyph_index() = glyph;
461 	  _hb_glyph_info_set_unicode_props (&buffer->out_info[starter], buffer);
462 
463 	  continue;
464 	}
465       }
466 
467       /* Blocked, or doesn't compose. */
468       buffer->next_glyph ();
469 
470       if (info_cc (buffer->prev()) == 0)
471 	starter = buffer->out_len - 1;
472     }
473     buffer->swap_buffers ();
474   }
475 }
476 
477 
478 #endif
479