• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // © 2016 and later: Unicode, Inc. and others.
2 // License & terms of use: http://www.unicode.org/copyright.html#License
3 /*
4 *******************************************************************************
5 * Copyright (C) 2013-2015, International Business Machines
6 * Corporation and others.  All Rights Reserved.
7 *******************************************************************************
8 * CollationFastLatinBuilder.java, ported from collationfastlatinbuilder.h/.cpp
9 *
10 * C++ version created on: 2013aug09
11 * created by: Markus W. Scherer
12 */
13 
14 package com.ibm.icu.impl.coll;
15 
16 import com.ibm.icu.lang.UScript;
17 import com.ibm.icu.text.Collator;
18 import com.ibm.icu.util.CharsTrie;
19 
20 final class CollationFastLatinBuilder {
21     // #define DEBUG_COLLATION_FAST_LATIN_BUILDER 0  // 0 or 1 or 2
22 
23     /**
24      * Compare two signed long values as if they were unsigned.
25      */
compareInt64AsUnsigned(long a, long b)26     private static final int compareInt64AsUnsigned(long a, long b) {
27         a += 0x8000000000000000L;
28         b += 0x8000000000000000L;
29         if(a < b) {
30             return -1;
31         } else if(a > b) {
32             return 1;
33         } else {
34             return 0;
35         }
36     }
37 
38     /**
39      * Like Java Collections.binarySearch(List, String, Comparator).
40      *
41      * @return the index>=0 where the item was found,
42      *         or the index<0 for inserting the string at ~index in sorted order
43      */
binarySearch(long[] list, int limit, long ce)44     private static final int binarySearch(long[] list, int limit, long ce) {
45         if (limit == 0) { return ~0; }
46         int start = 0;
47         for (;;) {
48             int i = (int)(((long)start + (long)limit) / 2);
49             int cmp = compareInt64AsUnsigned(ce, list[i]);
50             if (cmp == 0) {
51                 return i;
52             } else if (cmp < 0) {
53                 if (i == start) {
54                     return ~start;  // insert ce before i
55                 }
56                 limit = i;
57             } else {
58                 if (i == start) {
59                     return ~(start + 1);  // insert ce after i
60                 }
61                 start = i;
62             }
63         }
64     }
65 
CollationFastLatinBuilder()66     CollationFastLatinBuilder() {
67         ce0 = 0;
68         ce1 = 0;
69         contractionCEs = new UVector64();
70         uniqueCEs = new UVector64();
71         miniCEs = null;
72         firstDigitPrimary = 0;
73         firstLatinPrimary = 0;
74         lastLatinPrimary = 0;
75         firstShortPrimary = 0;
76         shortPrimaryOverflow = false;
77         headerLength = 0;
78     }
79 
forData(CollationData data)80     boolean forData(CollationData data) {
81         if(result.length() != 0) {  // This builder is not reusable.
82             throw new IllegalStateException("attempt to reuse a CollationFastLatinBuilder");
83         }
84         if(!loadGroups(data)) { return false; }
85 
86         // Fast handling of digits.
87         firstShortPrimary = firstDigitPrimary;
88         getCEs(data);
89         encodeUniqueCEs();
90         if(shortPrimaryOverflow) {
91             // Give digits long mini primaries,
92             // so that there are more short primaries for letters.
93             firstShortPrimary = firstLatinPrimary;
94             resetCEs();
95             getCEs(data);
96             encodeUniqueCEs();
97         }
98         // Note: If we still have a short-primary overflow but not a long-primary overflow,
99         // then we could calculate how many more long primaries would fit,
100         // and set the firstShortPrimary to that many after the current firstShortPrimary,
101         // and try again.
102         // However, this might only benefit the en_US_POSIX tailoring,
103         // and it is simpler to suppress building fast Latin data for it in genrb,
104         // or by returning false here if shortPrimaryOverflow.
105 
106         boolean ok = !shortPrimaryOverflow;
107         if(ok) {
108             encodeCharCEs();
109             encodeContractions();
110         }
111         contractionCEs.removeAllElements();  // might reduce heap memory usage
112         uniqueCEs.removeAllElements();
113         return ok;
114     }
115 
116     // C++ returns one combined array with the contents of the result buffer.
117     // Java returns two arrays (header & table) because we cannot use pointer arithmetic,
118     // and we do not want to index into the table with an offset.
getHeader()119     char[] getHeader() {
120         char[] resultArray = new char[headerLength];
121         result.getChars(0, headerLength, resultArray, 0);
122         return resultArray;
123     }
124 
getTable()125     char[] getTable() {
126         char[] resultArray = new char[result.length() - headerLength];
127         result.getChars(headerLength, result.length(), resultArray, 0);
128         return resultArray;
129     }
130 
loadGroups(CollationData data)131     private boolean loadGroups(CollationData data) {
132         headerLength = 1 + NUM_SPECIAL_GROUPS;
133         int r0 = (CollationFastLatin.VERSION << 8) | headerLength;
134         result.append((char)r0);
135         // The first few reordering groups should be special groups
136         // (space, punct, ..., digit) followed by Latn, then Grek and other scripts.
137         for(int i = 0; i < NUM_SPECIAL_GROUPS; ++i) {
138             lastSpecialPrimaries[i] = data.getLastPrimaryForGroup(Collator.ReorderCodes.FIRST + i);
139             if(lastSpecialPrimaries[i] == 0) {
140                 // missing data
141                 return false;
142             }
143             result.append(0);  // reserve a slot for this group
144         }
145 
146         firstDigitPrimary = data.getFirstPrimaryForGroup(Collator.ReorderCodes.DIGIT);
147         firstLatinPrimary = data.getFirstPrimaryForGroup(UScript.LATIN);
148         lastLatinPrimary = data.getLastPrimaryForGroup(UScript.LATIN);
149         if(firstDigitPrimary == 0 || firstLatinPrimary == 0) {
150             // missing data
151             return false;
152         }
153         return true;
154     }
155 
inSameGroup(long p, long q)156     private boolean inSameGroup(long p, long q) {
157         // Both or neither need to be encoded as short primaries,
158         // so that we can test only one and use the same bit mask.
159         if(p >= firstShortPrimary) {
160             return q >= firstShortPrimary;
161         } else if(q >= firstShortPrimary) {
162             return false;
163         }
164         // Both or neither must be potentially-variable,
165         // so that we can test only one and determine if both are variable.
166         long lastVariablePrimary = lastSpecialPrimaries[NUM_SPECIAL_GROUPS - 1];
167         if(p > lastVariablePrimary) {
168             return q > lastVariablePrimary;
169         } else if(q > lastVariablePrimary) {
170             return false;
171         }
172         // Both will be encoded with long mini primaries.
173         // They must be in the same special reordering group,
174         // so that we can test only one and determine if both are variable.
175         assert(p != 0 && q != 0);
176         for(int i = 0;; ++i) {  // will terminate
177             long lastPrimary = lastSpecialPrimaries[i];
178             if(p <= lastPrimary) {
179                 return q <= lastPrimary;
180             } else if(q <= lastPrimary) {
181                 return false;
182             }
183         }
184     }
185 
resetCEs()186     private void resetCEs() {
187         contractionCEs.removeAllElements();
188         uniqueCEs.removeAllElements();
189         shortPrimaryOverflow = false;
190         result.setLength(headerLength);
191     }
192 
getCEs(CollationData data)193     private void getCEs(CollationData data) {
194         int i = 0;
195         for(char c = 0;; ++i, ++c) {
196             if(c == CollationFastLatin.LATIN_LIMIT) {
197                 c = CollationFastLatin.PUNCT_START;
198             } else if(c == CollationFastLatin.PUNCT_LIMIT) {
199                 break;
200             }
201             CollationData d;
202             int ce32 = data.getCE32(c);
203             if(ce32 == Collation.FALLBACK_CE32) {
204                 d = data.base;
205                 ce32 = d.getCE32(c);
206             } else {
207                 d = data;
208             }
209             if(getCEsFromCE32(d, c, ce32)) {
210                 charCEs[i][0] = ce0;
211                 charCEs[i][1] = ce1;
212                 addUniqueCE(ce0);
213                 addUniqueCE(ce1);
214             } else {
215                 // bail out for c
216                 charCEs[i][0] = ce0 = Collation.NO_CE;
217                 charCEs[i][1] = ce1 = 0;
218             }
219             if(c == 0 && !isContractionCharCE(ce0)) {
220                 // Always map U+0000 to a contraction.
221                 // Write a contraction list with only a default value if there is no real contraction.
222                 assert(contractionCEs.isEmpty());
223                 addContractionEntry(CollationFastLatin.CONTR_CHAR_MASK, ce0, ce1);
224                 charCEs[0][0] = (Collation.NO_CE_PRIMARY << 32) | CONTRACTION_FLAG;
225                 charCEs[0][1] = 0;
226             }
227         }
228         // Terminate the last contraction list.
229         contractionCEs.addElement(CollationFastLatin.CONTR_CHAR_MASK);
230     }
231 
getCEsFromCE32(CollationData data, int c, int ce32)232     private boolean getCEsFromCE32(CollationData data, int c, int ce32) {
233         ce32 = data.getFinalCE32(ce32);
234         ce1 = 0;
235         if(Collation.isSimpleOrLongCE32(ce32)) {
236             ce0 = Collation.ceFromCE32(ce32);
237         } else {
238             switch(Collation.tagFromCE32(ce32)) {
239             case Collation.LATIN_EXPANSION_TAG:
240                 ce0 = Collation.latinCE0FromCE32(ce32);
241                 ce1 = Collation.latinCE1FromCE32(ce32);
242                 break;
243             case Collation.EXPANSION32_TAG: {
244                 int index = Collation.indexFromCE32(ce32);
245                 int length = Collation.lengthFromCE32(ce32);
246                 if(length <= 2) {
247                     ce0 = Collation.ceFromCE32(data.ce32s[index]);
248                     if(length == 2) {
249                         ce1 = Collation.ceFromCE32(data.ce32s[index + 1]);
250                     }
251                     break;
252                 } else {
253                     return false;
254                 }
255             }
256             case Collation.EXPANSION_TAG: {
257                 int index = Collation.indexFromCE32(ce32);
258                 int length = Collation.lengthFromCE32(ce32);
259                 if(length <= 2) {
260                     ce0 = data.ces[index];
261                     if(length == 2) {
262                         ce1 = data.ces[index + 1];
263                     }
264                     break;
265                 } else {
266                     return false;
267                 }
268             }
269             // Note: We could support PREFIX_TAG (assert c>=0)
270             // by recursing on its default CE32 and checking that none of the prefixes starts
271             // with a fast Latin character.
272             // However, currently (2013) there are only the L-before-middle-dot
273             // prefix mappings in the Latin range, and those would be rejected anyway.
274             case Collation.CONTRACTION_TAG:
275                 assert(c >= 0);
276                 return getCEsFromContractionCE32(data, ce32);
277             case Collation.OFFSET_TAG:
278                 assert(c >= 0);
279                 ce0 = data.getCEFromOffsetCE32(c, ce32);
280                 break;
281             default:
282                 return false;
283             }
284         }
285         // A mapping can be completely ignorable.
286         if(ce0 == 0) { return ce1 == 0; }
287         // We do not support an ignorable ce0 unless it is completely ignorable.
288         long p0 = ce0 >>> 32;
289         if(p0 == 0) { return false; }
290         // We only support primaries up to the Latin script.
291         if(p0 > lastLatinPrimary) { return false; }
292         // We support non-common secondary and case weights only together with short primaries.
293         int lower32_0 = (int)ce0;
294         if(p0 < firstShortPrimary) {
295             int sc0 = lower32_0 & Collation.SECONDARY_AND_CASE_MASK;
296             if(sc0 != Collation.COMMON_SECONDARY_CE) { return false; }
297         }
298         // No below-common tertiary weights.
299         if((lower32_0 & Collation.ONLY_TERTIARY_MASK) < Collation.COMMON_WEIGHT16) { return false; }
300         if(ce1 != 0) {
301             // Both primaries must be in the same group,
302             // or both must get short mini primaries,
303             // or a short-primary CE is followed by a secondary CE.
304             // This is so that we can test the first primary and use the same mask for both,
305             // and determine for both whether they are variable.
306             long p1 = ce1 >>> 32;
307             if(p1 == 0 ? p0 < firstShortPrimary : !inSameGroup(p0, p1)) { return false; }
308             int lower32_1 = (int)ce1;
309             // No tertiary CEs.
310             if((lower32_1 >>> 16) == 0) { return false; }
311             // We support non-common secondary and case weights
312             // only for secondary CEs or together with short primaries.
313             if(p1 != 0 && p1 < firstShortPrimary) {
314                 int sc1 = lower32_1 & Collation.SECONDARY_AND_CASE_MASK;
315                 if(sc1 != Collation.COMMON_SECONDARY_CE) { return false; }
316             }
317             // No below-common tertiary weights.
318             if((lower32_0 & Collation.ONLY_TERTIARY_MASK) < Collation.COMMON_WEIGHT16) { return false; }
319         }
320         // No quaternary weights.
321         if(((ce0 | ce1) & Collation.QUATERNARY_MASK) != 0) { return false; }
322         return true;
323     }
324 
getCEsFromContractionCE32(CollationData data, int ce32)325     private boolean getCEsFromContractionCE32(CollationData data, int ce32) {
326         int trieIndex = Collation.indexFromCE32(ce32);
327         ce32 = data.getCE32FromContexts(trieIndex);  // Default if no suffix match.
328         // Since the original ce32 is not a prefix mapping,
329         // the default ce32 must not be another contraction.
330         assert(!Collation.isContractionCE32(ce32));
331         int contractionIndex = contractionCEs.size();
332         if(getCEsFromCE32(data, Collation.SENTINEL_CP, ce32)) {
333             addContractionEntry(CollationFastLatin.CONTR_CHAR_MASK, ce0, ce1);
334         } else {
335             // Bail out for c-without-contraction.
336             addContractionEntry(CollationFastLatin.CONTR_CHAR_MASK, Collation.NO_CE, 0);
337         }
338         // Handle an encodable contraction unless the next contraction is too long
339         // and starts with the same character.
340         int prevX = -1;
341         boolean addContraction = false;
342         CharsTrie.Iterator suffixes = CharsTrie.iterator(data.contexts, trieIndex + 2, 0);
343         while(suffixes.hasNext()) {
344             CharsTrie.Entry entry = suffixes.next();
345             CharSequence suffix = entry.chars;
346             int x = CollationFastLatin.getCharIndex(suffix.charAt(0));
347             if(x < 0) { continue; }  // ignore anything but fast Latin text
348             if(x == prevX) {
349                 if(addContraction) {
350                     // Bail out for all contractions starting with this character.
351                     addContractionEntry(x, Collation.NO_CE, 0);
352                     addContraction = false;
353                 }
354                 continue;
355             }
356             if(addContraction) {
357                 addContractionEntry(prevX, ce0, ce1);
358             }
359             ce32 = entry.value;
360             if(suffix.length() == 1 && getCEsFromCE32(data, Collation.SENTINEL_CP, ce32)) {
361                 addContraction = true;
362             } else {
363                 addContractionEntry(x, Collation.NO_CE, 0);
364                 addContraction = false;
365             }
366             prevX = x;
367         }
368         if(addContraction) {
369             addContractionEntry(prevX, ce0, ce1);
370         }
371         // Note: There might not be any fast Latin contractions, but
372         // we need to enter contraction handling anyway so that we can bail out
373         // when there is a non-fast-Latin character following.
374         // For example: Danish &Y<<u+umlaut, when we compare Y vs. u\u0308 we need to see the
375         // following umlaut and bail out, rather than return the difference of Y vs. u.
376         ce0 = (Collation.NO_CE_PRIMARY << 32) | CONTRACTION_FLAG | contractionIndex;
377         ce1 = 0;
378         return true;
379     }
380 
addContractionEntry(int x, long cce0, long cce1)381     private void addContractionEntry(int x, long cce0, long cce1) {
382         contractionCEs.addElement(x);
383         contractionCEs.addElement(cce0);
384         contractionCEs.addElement(cce1);
385         addUniqueCE(cce0);
386         addUniqueCE(cce1);
387     }
388 
addUniqueCE(long ce)389     private void addUniqueCE(long ce) {
390         if(ce == 0 || (ce >>> 32) == Collation.NO_CE_PRIMARY) { return; }
391         ce &= ~(long)Collation.CASE_MASK;  // blank out case bits
392         int i = binarySearch(uniqueCEs.getBuffer(), uniqueCEs.size(), ce);
393         if(i < 0) {
394             uniqueCEs.insertElementAt(ce, ~i);
395         }
396     }
397 
getMiniCE(long ce)398     private int getMiniCE(long ce) {
399         ce &= ~(long)Collation.CASE_MASK;  // blank out case bits
400         int index = binarySearch(uniqueCEs.getBuffer(), uniqueCEs.size(), ce);
401         assert(index >= 0);
402         return miniCEs[index];
403     }
404 
encodeUniqueCEs()405     private void encodeUniqueCEs() {
406         miniCEs = new char[uniqueCEs.size()];
407         int group = 0;
408         long lastGroupPrimary = lastSpecialPrimaries[group];
409         // The lowest unique CE must be at least a secondary CE.
410         assert(((int)uniqueCEs.elementAti(0) >>> 16) != 0);
411         long prevPrimary = 0;
412         int prevSecondary = 0;
413         int pri = 0;
414         int sec = 0;
415         int ter = CollationFastLatin.COMMON_TER;
416         for(int i = 0; i < uniqueCEs.size(); ++i) {
417             long ce = uniqueCEs.elementAti(i);
418             // Note: At least one of the p/s/t weights changes from one unique CE to the next.
419             // (uniqueCEs does not store case bits.)
420             long p = ce >>> 32;
421             if(p != prevPrimary) {
422                 while(p > lastGroupPrimary) {
423                     assert(pri <= CollationFastLatin.MAX_LONG);
424                     // Set the group's header entry to the
425                     // last "long primary" in or before the group.
426                     result.setCharAt(1 + group, (char)pri);
427                     if(++group < NUM_SPECIAL_GROUPS) {
428                         lastGroupPrimary = lastSpecialPrimaries[group];
429                     } else {
430                         lastGroupPrimary = 0xffffffffL;
431                         break;
432                     }
433                 }
434                 if(p < firstShortPrimary) {
435                     if(pri == 0) {
436                         pri = CollationFastLatin.MIN_LONG;
437                     } else if(pri < CollationFastLatin.MAX_LONG) {
438                         pri += CollationFastLatin.LONG_INC;
439                     } else {
440     /* #if DEBUG_COLLATION_FAST_LATIN_BUILDER
441                         printf("long-primary overflow for %08x\n", p);
442     #endif */
443                         miniCEs[i] = CollationFastLatin.BAIL_OUT;
444                         continue;
445                     }
446                 } else {
447                     if(pri < CollationFastLatin.MIN_SHORT) {
448                         pri = CollationFastLatin.MIN_SHORT;
449                     } else if(pri < (CollationFastLatin.MAX_SHORT - CollationFastLatin.SHORT_INC)) {
450                         // Reserve the highest primary weight for U+FFFF.
451                         pri += CollationFastLatin.SHORT_INC;
452                     } else {
453     /* #if DEBUG_COLLATION_FAST_LATIN_BUILDER
454                         printf("short-primary overflow for %08x\n", p);
455     #endif */
456                         shortPrimaryOverflow = true;
457                         miniCEs[i] = CollationFastLatin.BAIL_OUT;
458                         continue;
459                     }
460                 }
461                 prevPrimary = p;
462                 prevSecondary = Collation.COMMON_WEIGHT16;
463                 sec = CollationFastLatin.COMMON_SEC;
464                 ter = CollationFastLatin.COMMON_TER;
465             }
466             int lower32 = (int)ce;
467             int s = lower32 >>> 16;
468             if(s != prevSecondary) {
469                 if(pri == 0) {
470                     if(sec == 0) {
471                         sec = CollationFastLatin.MIN_SEC_HIGH;
472                     } else if(sec < CollationFastLatin.MAX_SEC_HIGH) {
473                         sec += CollationFastLatin.SEC_INC;
474                     } else {
475                         miniCEs[i] = CollationFastLatin.BAIL_OUT;
476                         continue;
477                     }
478                     prevSecondary = s;
479                     ter = CollationFastLatin.COMMON_TER;
480                 } else if(s < Collation.COMMON_WEIGHT16) {
481                     if(sec == CollationFastLatin.COMMON_SEC) {
482                         sec = CollationFastLatin.MIN_SEC_BEFORE;
483                     } else if(sec < CollationFastLatin.MAX_SEC_BEFORE) {
484                         sec += CollationFastLatin.SEC_INC;
485                     } else {
486                         miniCEs[i] = CollationFastLatin.BAIL_OUT;
487                         continue;
488                     }
489                 } else if(s == Collation.COMMON_WEIGHT16) {
490                     sec = CollationFastLatin.COMMON_SEC;
491                 } else {
492                     if(sec < CollationFastLatin.MIN_SEC_AFTER) {
493                         sec = CollationFastLatin.MIN_SEC_AFTER;
494                     } else if(sec < CollationFastLatin.MAX_SEC_AFTER) {
495                         sec += CollationFastLatin.SEC_INC;
496                     } else {
497                         miniCEs[i] = CollationFastLatin.BAIL_OUT;
498                         continue;
499                     }
500                 }
501                 prevSecondary = s;
502                 ter = CollationFastLatin.COMMON_TER;
503             }
504             assert((lower32 & Collation.CASE_MASK) == 0);  // blanked out in uniqueCEs
505             int t = lower32 & Collation.ONLY_TERTIARY_MASK;
506             if(t > Collation.COMMON_WEIGHT16) {
507                 if(ter < CollationFastLatin.MAX_TER_AFTER) {
508                     ++ter;
509                 } else {
510                     miniCEs[i] = CollationFastLatin.BAIL_OUT;
511                     continue;
512                 }
513             }
514             if(CollationFastLatin.MIN_LONG <= pri && pri <= CollationFastLatin.MAX_LONG) {
515                 assert(sec == CollationFastLatin.COMMON_SEC);
516                 miniCEs[i] = (char)(pri | ter);
517             } else {
518                 miniCEs[i] = (char)(pri | sec | ter);
519             }
520         }
521     /* #if DEBUG_COLLATION_FAST_LATIN_BUILDER
522         printf("last mini primary: %04x\n", pri);
523     #endif */
524     /* #if DEBUG_COLLATION_FAST_LATIN_BUILDER >= 2
525         for(int i = 0; i < uniqueCEs.size(); ++i) {
526             long ce = uniqueCEs.elementAti(i);
527             printf("unique CE 0x%016lx -> 0x%04x\n", ce, miniCEs[i]);
528         }
529     #endif */
530     }
531 
encodeCharCEs()532     private void encodeCharCEs() {
533         int miniCEsStart = result.length();
534         for(int i = 0; i < CollationFastLatin.NUM_FAST_CHARS; ++i) {
535             result.append(0);  // initialize to completely ignorable
536         }
537         int indexBase = result.length();
538         for(int i = 0; i < CollationFastLatin.NUM_FAST_CHARS; ++i) {
539             long ce = charCEs[i][0];
540             if(isContractionCharCE(ce)) { continue; }  // defer contraction
541             int miniCE = encodeTwoCEs(ce, charCEs[i][1]);
542             if((miniCE >>> 16) > 0) {   // if ((unsigned)miniCE > 0xffff)
543                 // Note: There is a chance that this new expansion is the same as a previous one,
544                 // and if so, then we could reuse the other expansion.
545                 // However, that seems unlikely.
546                 int expansionIndex = result.length() - indexBase;
547                 if(expansionIndex > CollationFastLatin.INDEX_MASK) {
548                     miniCE = CollationFastLatin.BAIL_OUT;
549                 } else {
550                     result.append((char)(miniCE >> 16)).append((char)miniCE);
551                     miniCE = CollationFastLatin.EXPANSION | expansionIndex;
552                 }
553             }
554             result.setCharAt(miniCEsStart + i, (char)miniCE);
555         }
556     }
557 
encodeContractions()558     private void encodeContractions() {
559         // We encode all contraction lists so that the first word of a list
560         // terminates the previous list, and we only need one additional terminator at the end.
561         int indexBase = headerLength + CollationFastLatin.NUM_FAST_CHARS;
562         int firstContractionIndex = result.length();
563         for(int i = 0; i < CollationFastLatin.NUM_FAST_CHARS; ++i) {
564             long ce = charCEs[i][0];
565             if(!isContractionCharCE(ce)) { continue; }
566             int contractionIndex = result.length() - indexBase;
567             if(contractionIndex > CollationFastLatin.INDEX_MASK) {
568                 result.setCharAt(headerLength + i, (char) CollationFastLatin.BAIL_OUT);
569                 continue;
570             }
571             boolean firstTriple = true;
572             for(int index = (int)ce & 0x7fffffff;; index += 3) {
573                 long x = contractionCEs.elementAti(index);
574                 if(x == CollationFastLatin.CONTR_CHAR_MASK && !firstTriple) { break; }
575                 long cce0 = contractionCEs.elementAti(index + 1);
576                 long cce1 = contractionCEs.elementAti(index + 2);
577                 int miniCE = encodeTwoCEs(cce0, cce1);
578                 if(miniCE == CollationFastLatin.BAIL_OUT) {
579                     result.append((char)(x | (1 << CollationFastLatin.CONTR_LENGTH_SHIFT)));
580                 } else if((miniCE >>> 16) == 0) {  // if ((unsigned)miniCE <= 0xffff)
581                     result.append((char)(x | (2 << CollationFastLatin.CONTR_LENGTH_SHIFT)));
582                     result.append((char)miniCE);
583                 } else {
584                     result.append((char)(x | (3 << CollationFastLatin.CONTR_LENGTH_SHIFT)));
585                     result.append((char)(miniCE >> 16)).append((char)miniCE);
586                 }
587                 firstTriple = false;
588             }
589             // Note: There is a chance that this new contraction list is the same as a previous one,
590             // and if so, then we could truncate the result and reuse the other list.
591             // However, that seems unlikely.
592             result.setCharAt(headerLength + i,
593                             (char)(CollationFastLatin.CONTRACTION | contractionIndex));
594         }
595         if(result.length() > firstContractionIndex) {
596             // Terminate the last contraction list.
597             result.append((char)CollationFastLatin.CONTR_CHAR_MASK);
598         }
599     /* #if DEBUG_COLLATION_FAST_LATIN_BUILDER
600         printf("** fast Latin %d * 2 = %d bytes\n", result.length(), result.length() * 2);
601         puts("   header & below-digit groups map");
602         int i = 0;
603         for(; i < headerLength; ++i) {
604             printf(" %04x", result[i]);
605         }
606         printf("\n   char mini CEs");
607         assert(CollationFastLatin.NUM_FAST_CHARS % 16 == 0);
608         for(; i < indexBase; i += 16) {
609             int c = i - headerLength;
610             if(c >= CollationFastLatin.LATIN_LIMIT) {
611                 c = CollationFastLatin.PUNCT_START + c - CollationFastLatin.LATIN_LIMIT;
612             }
613             printf("\n %04x:", c);
614             for(int j = 0; j < 16; ++j) {
615                 printf(" %04x", result[i + j]);
616             }
617         }
618         printf("\n   expansions & contractions");
619         for(; i < result.length(); ++i) {
620             if((i - indexBase) % 16 == 0) { puts(""); }
621             printf(" %04x", result[i]);
622         }
623         puts("");
624     #endif */
625     }
626 
encodeTwoCEs(long first, long second)627     private int encodeTwoCEs(long first, long second) {
628         if(first == 0) {
629             return 0;  // completely ignorable
630         }
631         if(first == Collation.NO_CE) {
632             return CollationFastLatin.BAIL_OUT;
633         }
634         assert((first >>> 32) != Collation.NO_CE_PRIMARY);
635 
636         int miniCE = getMiniCE(first);
637         if(miniCE == CollationFastLatin.BAIL_OUT) { return miniCE; }
638         if(miniCE >= CollationFastLatin.MIN_SHORT) {
639             // Extract & copy the case bits.
640             // Shift them from normal CE bits 15..14 to mini CE bits 4..3.
641             int c = (((int)first & Collation.CASE_MASK) >> (14 - 3));
642             // Only in mini CEs: Ignorable case bits = 0, lowercase = 1.
643             c += CollationFastLatin.LOWER_CASE;
644             miniCE |= c;
645         }
646         if(second == 0) { return miniCE; }
647 
648         int miniCE1 = getMiniCE(second);
649         if(miniCE1 == CollationFastLatin.BAIL_OUT) { return miniCE1; }
650 
651         int case1 = (int)second & Collation.CASE_MASK;
652         if(miniCE >= CollationFastLatin.MIN_SHORT &&
653                 (miniCE & CollationFastLatin.SECONDARY_MASK) == CollationFastLatin.COMMON_SEC) {
654             // Try to combine the two mini CEs into one.
655             int sec1 = miniCE1 & CollationFastLatin.SECONDARY_MASK;
656             int ter1 = miniCE1 & CollationFastLatin.TERTIARY_MASK;
657             if(sec1 >= CollationFastLatin.MIN_SEC_HIGH && case1 == 0 &&
658                     ter1 == CollationFastLatin.COMMON_TER) {
659                 // sec1>=sec_high implies pri1==0.
660                 return (miniCE & ~CollationFastLatin.SECONDARY_MASK) | sec1;
661             }
662         }
663 
664         if(miniCE1 <= CollationFastLatin.SECONDARY_MASK || CollationFastLatin.MIN_SHORT <= miniCE1) {
665             // Secondary CE, or a CE with a short primary, copy the case bits.
666             case1 = (case1 >> (14 - 3)) + CollationFastLatin.LOWER_CASE;
667             miniCE1 |= case1;
668         }
669         return (miniCE << 16) | miniCE1;
670     }
671 
isContractionCharCE(long ce)672     private static boolean isContractionCharCE(long ce) {
673         return (ce >>> 32) == Collation.NO_CE_PRIMARY && ce != Collation.NO_CE;
674     }
675 
676     // space, punct, symbol, currency (not digit)
677     private static final int NUM_SPECIAL_GROUPS =
678             Collator.ReorderCodes.CURRENCY - Collator.ReorderCodes.FIRST + 1;
679 
680     private static final long CONTRACTION_FLAG = 0x80000000L;
681 
682     // temporary "buffer"
683     private long ce0, ce1;
684 
685     private long[][] charCEs = new long[CollationFastLatin.NUM_FAST_CHARS][2];
686 
687     private UVector64 contractionCEs;
688     private UVector64 uniqueCEs;
689 
690     /** One 16-bit mini CE per unique CE. */
691     private char[] miniCEs;
692 
693     // These are constant for a given root collator.
694     long[] lastSpecialPrimaries = new long[NUM_SPECIAL_GROUPS];
695     private long firstDigitPrimary;
696     private long firstLatinPrimary;
697     private long lastLatinPrimary;
698     // This determines the first normal primary weight which is mapped to
699     // a short mini primary. It must be >=firstDigitPrimary.
700     private long firstShortPrimary;
701 
702     private boolean shortPrimaryOverflow;
703 
704     private StringBuilder result = new StringBuilder();
705     private int headerLength;
706 }
707