1 /* 2 ******************************************************************************* 3 * Copyright (C) 2013-2014, International Business Machines 4 * Corporation and others. All Rights Reserved. 5 ******************************************************************************* 6 * collationfastlatinbuilder.cpp 7 * 8 * created on: 2013aug09 9 * created by: Markus W. Scherer 10 */ 11 12 #define DEBUG_COLLATION_FAST_LATIN_BUILDER 0 // 0 or 1 or 2 13 #if DEBUG_COLLATION_FAST_LATIN_BUILDER 14 #include <stdio.h> 15 #include <string> 16 #endif 17 18 #include "unicode/utypes.h" 19 20 #if !UCONFIG_NO_COLLATION 21 22 #include "unicode/ucol.h" 23 #include "unicode/ucharstrie.h" 24 #include "unicode/unistr.h" 25 #include "unicode/uobject.h" 26 #include "unicode/uscript.h" 27 #include "cmemory.h" 28 #include "collation.h" 29 #include "collationdata.h" 30 #include "collationfastlatin.h" 31 #include "collationfastlatinbuilder.h" 32 #include "uassert.h" 33 #include "uvectr64.h" 34 35 U_NAMESPACE_BEGIN 36 37 struct CollationData; 38 39 namespace { 40 41 /** 42 * Compare two signed int64_t values as if they were unsigned. 43 */ 44 int32_t 45 compareInt64AsUnsigned(int64_t a, int64_t b) { 46 if((uint64_t)a < (uint64_t)b) { 47 return -1; 48 } else if((uint64_t)a > (uint64_t)b) { 49 return 1; 50 } else { 51 return 0; 52 } 53 } 54 55 // TODO: Merge this with the near-identical version in collationbasedatabuilder.cpp 56 /** 57 * Like Java Collections.binarySearch(List, String, Comparator). 58 * 59 * @return the index>=0 where the item was found, 60 * or the index<0 for inserting the string at ~index in sorted order 61 */ 62 int32_t 63 binarySearch(const int64_t list[], int32_t limit, int64_t ce) { 64 if (limit == 0) { return ~0; } 65 int32_t start = 0; 66 for (;;) { 67 int32_t i = (start + limit) / 2; 68 int32_t cmp = compareInt64AsUnsigned(ce, list[i]); 69 if (cmp == 0) { 70 return i; 71 } else if (cmp < 0) { 72 if (i == start) { 73 return ~start; // insert ce before i 74 } 75 limit = i; 76 } else { 77 if (i == start) { 78 return ~(start + 1); // insert ce after i 79 } 80 start = i; 81 } 82 } 83 } 84 85 } // namespace 86 87 CollationFastLatinBuilder::CollationFastLatinBuilder(UErrorCode &errorCode) 88 : ce0(0), ce1(0), 89 contractionCEs(errorCode), uniqueCEs(errorCode), 90 miniCEs(NULL), 91 firstDigitPrimary(0), firstLatinPrimary(0), lastLatinPrimary(0), 92 firstShortPrimary(0), shortPrimaryOverflow(FALSE), 93 headerLength(0) { 94 } 95 96 CollationFastLatinBuilder::~CollationFastLatinBuilder() { 97 uprv_free(miniCEs); 98 } 99 100 UBool 101 CollationFastLatinBuilder::forData(const CollationData &data, UErrorCode &errorCode) { 102 if(U_FAILURE(errorCode)) { return FALSE; } 103 if(!result.isEmpty()) { // This builder is not reusable. 104 errorCode = U_INVALID_STATE_ERROR; 105 return FALSE; 106 } 107 if(!loadGroups(data, errorCode)) { return FALSE; } 108 109 // Fast handling of digits. 110 firstShortPrimary = firstDigitPrimary; 111 getCEs(data, errorCode); 112 if(!encodeUniqueCEs(errorCode)) { return FALSE; } 113 if(shortPrimaryOverflow) { 114 // Give digits long mini primaries, 115 // so that there are more short primaries for letters. 116 firstShortPrimary = firstLatinPrimary; 117 resetCEs(); 118 getCEs(data, errorCode); 119 if(!encodeUniqueCEs(errorCode)) { return FALSE; } 120 } 121 // Note: If we still have a short-primary overflow but not a long-primary overflow, 122 // then we could calculate how many more long primaries would fit, 123 // and set the firstShortPrimary to that many after the current firstShortPrimary, 124 // and try again. 125 // However, this might only benefit the en_US_POSIX tailoring, 126 // and it is simpler to suppress building fast Latin data for it in genrb, 127 // or by returning FALSE here if shortPrimaryOverflow. 128 129 UBool ok = !shortPrimaryOverflow && 130 encodeCharCEs(errorCode) && encodeContractions(errorCode); 131 contractionCEs.removeAllElements(); // might reduce heap memory usage 132 uniqueCEs.removeAllElements(); 133 return ok; 134 } 135 136 UBool 137 CollationFastLatinBuilder::loadGroups(const CollationData &data, UErrorCode &errorCode) { 138 if(U_FAILURE(errorCode)) { return FALSE; } 139 result.append(0); // reserved for version & headerLength 140 // The first few reordering groups should be special groups 141 // (space, punct, ..., digit) followed by Latn, then Grek and other scripts. 142 for(int32_t i = 0;;) { 143 if(i >= data.scriptsLength) { 144 // no Latn script 145 errorCode = U_INTERNAL_PROGRAM_ERROR; 146 return FALSE; 147 } 148 uint32_t head = data.scripts[i]; 149 uint32_t lastByte = head & 0xff; // last primary byte in the group 150 int32_t group = data.scripts[i + 2]; 151 if(group == UCOL_REORDER_CODE_DIGIT) { 152 firstDigitPrimary = (head & 0xff00) << 16; 153 headerLength = result.length(); 154 uint32_t r0 = (CollationFastLatin::VERSION << 8) | headerLength; 155 result.setCharAt(0, (UChar)r0); 156 } else if(group == USCRIPT_LATIN) { 157 if(firstDigitPrimary == 0) { 158 // no digit group 159 errorCode = U_INTERNAL_PROGRAM_ERROR; 160 return FALSE; 161 } 162 firstLatinPrimary = (head & 0xff00) << 16; 163 lastLatinPrimary = (lastByte << 24) | 0xffffff; 164 break; 165 } else if(firstDigitPrimary == 0) { 166 // a group below digits 167 if(lastByte > 0x7f) { 168 // We only use 7 bits for the last byte of a below-digits group. 169 // This does not warrant an errorCode, but we do not build a fast Latin table. 170 return FALSE; 171 } 172 result.append((UChar)lastByte); 173 } 174 i = i + 2 + data.scripts[i + 1]; 175 } 176 return TRUE; 177 } 178 179 UBool 180 CollationFastLatinBuilder::inSameGroup(uint32_t p, uint32_t q) const { 181 // Both or neither need to be encoded as short primaries, 182 // so that we can test only one and use the same bit mask. 183 if(p >= firstShortPrimary) { 184 return q >= firstShortPrimary; 185 } else if(q >= firstShortPrimary) { 186 return FALSE; 187 } 188 // Both or neither must be potentially-variable, 189 // so that we can test only one and determine if both are variable. 190 if(p >= firstDigitPrimary) { 191 return q >= firstDigitPrimary; 192 } else if(q >= firstDigitPrimary) { 193 return FALSE; 194 } 195 // Both will be encoded with long mini primaries. 196 // They must be in the same special reordering group, 197 // so that we can test only one and determine if both are variable. 198 p >>= 24; // first primary byte 199 q >>= 24; 200 U_ASSERT(p != 0 && q != 0); 201 U_ASSERT(p <= result[headerLength - 1]); // the loop will terminate 202 for(int32_t i = 1;; ++i) { 203 uint32_t lastByte = result[i]; 204 if(p <= lastByte) { 205 return q <= lastByte; 206 } else if(q <= lastByte) { 207 return FALSE; 208 } 209 } 210 } 211 212 void 213 CollationFastLatinBuilder::resetCEs() { 214 contractionCEs.removeAllElements(); 215 uniqueCEs.removeAllElements(); 216 shortPrimaryOverflow = FALSE; 217 result.truncate(headerLength); 218 } 219 220 void 221 CollationFastLatinBuilder::getCEs(const CollationData &data, UErrorCode &errorCode) { 222 if(U_FAILURE(errorCode)) { return; } 223 int32_t i = 0; 224 for(UChar c = 0;; ++i, ++c) { 225 if(c == CollationFastLatin::LATIN_LIMIT) { 226 c = CollationFastLatin::PUNCT_START; 227 } else if(c == CollationFastLatin::PUNCT_LIMIT) { 228 break; 229 } 230 const CollationData *d; 231 uint32_t ce32 = data.getCE32(c); 232 if(ce32 == Collation::FALLBACK_CE32) { 233 d = data.base; 234 ce32 = d->getCE32(c); 235 } else { 236 d = &data; 237 } 238 if(getCEsFromCE32(*d, c, ce32, errorCode)) { 239 charCEs[i][0] = ce0; 240 charCEs[i][1] = ce1; 241 addUniqueCE(ce0, errorCode); 242 addUniqueCE(ce1, errorCode); 243 } else { 244 // bail out for c 245 charCEs[i][0] = ce0 = Collation::NO_CE; 246 charCEs[i][1] = ce1 = 0; 247 } 248 if(c == 0 && !isContractionCharCE(ce0)) { 249 // Always map U+0000 to a contraction. 250 // Write a contraction list with only a default value if there is no real contraction. 251 U_ASSERT(contractionCEs.isEmpty()); 252 addContractionEntry(CollationFastLatin::CONTR_CHAR_MASK, ce0, ce1, errorCode); 253 charCEs[0][0] = ((int64_t)Collation::NO_CE_PRIMARY << 32) | CONTRACTION_FLAG; 254 charCEs[0][1] = 0; 255 } 256 } 257 // Terminate the last contraction list. 258 contractionCEs.addElement(CollationFastLatin::CONTR_CHAR_MASK, errorCode); 259 } 260 261 UBool 262 CollationFastLatinBuilder::getCEsFromCE32(const CollationData &data, UChar32 c, uint32_t ce32, 263 UErrorCode &errorCode) { 264 if(U_FAILURE(errorCode)) { return FALSE; } 265 ce32 = data.getFinalCE32(ce32); 266 ce1 = 0; 267 if(Collation::isSimpleOrLongCE32(ce32)) { 268 ce0 = Collation::ceFromCE32(ce32); 269 } else { 270 switch(Collation::tagFromCE32(ce32)) { 271 case Collation::LATIN_EXPANSION_TAG: 272 ce0 = Collation::latinCE0FromCE32(ce32); 273 ce1 = Collation::latinCE1FromCE32(ce32); 274 break; 275 case Collation::EXPANSION32_TAG: { 276 const uint32_t *ce32s = data.ce32s + Collation::indexFromCE32(ce32); 277 int32_t length = Collation::lengthFromCE32(ce32); 278 if(length <= 2) { 279 ce0 = Collation::ceFromCE32(ce32s[0]); 280 if(length == 2) { 281 ce1 = Collation::ceFromCE32(ce32s[1]); 282 } 283 break; 284 } else { 285 return FALSE; 286 } 287 } 288 case Collation::EXPANSION_TAG: { 289 const int64_t *ces = data.ces + Collation::indexFromCE32(ce32); 290 int32_t length = Collation::lengthFromCE32(ce32); 291 if(length <= 2) { 292 ce0 = ces[0]; 293 if(length == 2) { 294 ce1 = ces[1]; 295 } 296 break; 297 } else { 298 return FALSE; 299 } 300 } 301 // Note: We could support PREFIX_TAG (assert c>=0) 302 // by recursing on its default CE32 and checking that none of the prefixes starts 303 // with a fast Latin character. 304 // However, currently (2013) there are only the L-before-middle-dot 305 // prefix mappings in the Latin range, and those would be rejected anyway. 306 case Collation::CONTRACTION_TAG: 307 U_ASSERT(c >= 0); 308 return getCEsFromContractionCE32(data, ce32, errorCode); 309 case Collation::OFFSET_TAG: 310 U_ASSERT(c >= 0); 311 ce0 = data.getCEFromOffsetCE32(c, ce32); 312 break; 313 default: 314 return FALSE; 315 } 316 } 317 // A mapping can be completely ignorable. 318 if(ce0 == 0) { return ce1 == 0; } 319 // We do not support an ignorable ce0 unless it is completely ignorable. 320 uint32_t p0 = (uint32_t)(ce0 >> 32); 321 if(p0 == 0) { return FALSE; } 322 // We only support primaries up to the Latin script. 323 if(p0 > lastLatinPrimary) { return FALSE; } 324 // We support non-common secondary and case weights only together with short primaries. 325 uint32_t lower32_0 = (uint32_t)ce0; 326 if(p0 < firstShortPrimary) { 327 uint32_t sc0 = lower32_0 & Collation::SECONDARY_AND_CASE_MASK; 328 if(sc0 != Collation::COMMON_SECONDARY_CE) { return FALSE; } 329 } 330 // No below-common tertiary weights. 331 if((lower32_0 & Collation::ONLY_TERTIARY_MASK) < Collation::COMMON_WEIGHT16) { return FALSE; } 332 if(ce1 != 0) { 333 // Both primaries must be in the same group, 334 // or both must get short mini primaries, 335 // or a short-primary CE is followed by a secondary CE. 336 // This is so that we can test the first primary and use the same mask for both, 337 // and determine for both whether they are variable. 338 uint32_t p1 = (uint32_t)(ce1 >> 32); 339 if(p1 == 0 ? p0 < firstShortPrimary : !inSameGroup(p0, p1)) { return FALSE; } 340 uint32_t lower32_1 = (uint32_t)ce1; 341 // No tertiary CEs. 342 if((lower32_1 >> 16) == 0) { return FALSE; } 343 // We support non-common secondary and case weights 344 // only for secondary CEs or together with short primaries. 345 if(p1 != 0 && p1 < firstShortPrimary) { 346 uint32_t sc1 = lower32_1 & Collation::SECONDARY_AND_CASE_MASK; 347 if(sc1 != Collation::COMMON_SECONDARY_CE) { return FALSE; } 348 } 349 // No below-common tertiary weights. 350 if((lower32_1 & Collation::ONLY_TERTIARY_MASK) < Collation::COMMON_WEIGHT16) { return FALSE; } 351 } 352 // No quaternary weights. 353 if(((ce0 | ce1) & Collation::QUATERNARY_MASK) != 0) { return FALSE; } 354 return TRUE; 355 } 356 357 UBool 358 CollationFastLatinBuilder::getCEsFromContractionCE32(const CollationData &data, uint32_t ce32, 359 UErrorCode &errorCode) { 360 if(U_FAILURE(errorCode)) { return FALSE; } 361 const UChar *p = data.contexts + Collation::indexFromCE32(ce32); 362 ce32 = CollationData::readCE32(p); // Default if no suffix match. 363 // Since the original ce32 is not a prefix mapping, 364 // the default ce32 must not be another contraction. 365 U_ASSERT(!Collation::isContractionCE32(ce32)); 366 int32_t contractionIndex = contractionCEs.size(); 367 if(getCEsFromCE32(data, U_SENTINEL, ce32, errorCode)) { 368 addContractionEntry(CollationFastLatin::CONTR_CHAR_MASK, ce0, ce1, errorCode); 369 } else { 370 // Bail out for c-without-contraction. 371 addContractionEntry(CollationFastLatin::CONTR_CHAR_MASK, Collation::NO_CE, 0, errorCode); 372 } 373 // Handle an encodable contraction unless the next contraction is too long 374 // and starts with the same character. 375 int32_t prevX = -1; 376 UBool addContraction = FALSE; 377 UCharsTrie::Iterator suffixes(p + 2, 0, errorCode); 378 while(suffixes.next(errorCode)) { 379 const UnicodeString &suffix = suffixes.getString(); 380 int32_t x = CollationFastLatin::getCharIndex(suffix.charAt(0)); 381 if(x < 0) { continue; } // ignore anything but fast Latin text 382 if(x == prevX) { 383 if(addContraction) { 384 // Bail out for all contractions starting with this character. 385 addContractionEntry(x, Collation::NO_CE, 0, errorCode); 386 addContraction = FALSE; 387 } 388 continue; 389 } 390 if(addContraction) { 391 addContractionEntry(prevX, ce0, ce1, errorCode); 392 } 393 ce32 = (uint32_t)suffixes.getValue(); 394 if(suffix.length() == 1 && getCEsFromCE32(data, U_SENTINEL, ce32, errorCode)) { 395 addContraction = TRUE; 396 } else { 397 addContractionEntry(x, Collation::NO_CE, 0, errorCode); 398 addContraction = FALSE; 399 } 400 prevX = x; 401 } 402 if(addContraction) { 403 addContractionEntry(prevX, ce0, ce1, errorCode); 404 } 405 if(U_FAILURE(errorCode)) { return FALSE; } 406 // Note: There might not be any fast Latin contractions, but 407 // we need to enter contraction handling anyway so that we can bail out 408 // when there is a non-fast-Latin character following. 409 // For example: Danish &Y<<u+umlaut, when we compare Y vs. u\u0308 we need to see the 410 // following umlaut and bail out, rather than return the difference of Y vs. u. 411 ce0 = ((int64_t)Collation::NO_CE_PRIMARY << 32) | CONTRACTION_FLAG | contractionIndex; 412 ce1 = 0; 413 return TRUE; 414 } 415 416 void 417 CollationFastLatinBuilder::addContractionEntry(int32_t x, int64_t cce0, int64_t cce1, 418 UErrorCode &errorCode) { 419 contractionCEs.addElement(x, errorCode); 420 contractionCEs.addElement(cce0, errorCode); 421 contractionCEs.addElement(cce1, errorCode); 422 addUniqueCE(cce0, errorCode); 423 addUniqueCE(cce1, errorCode); 424 } 425 426 void 427 CollationFastLatinBuilder::addUniqueCE(int64_t ce, UErrorCode &errorCode) { 428 if(U_FAILURE(errorCode)) { return; } 429 if(ce == 0 || (uint32_t)(ce >> 32) == Collation::NO_CE_PRIMARY) { return; } 430 ce &= ~(int64_t)Collation::CASE_MASK; // blank out case bits 431 int32_t i = binarySearch(uniqueCEs.getBuffer(), uniqueCEs.size(), ce); 432 if(i < 0) { 433 uniqueCEs.insertElementAt(ce, ~i, errorCode); 434 } 435 } 436 437 uint32_t 438 CollationFastLatinBuilder::getMiniCE(int64_t ce) const { 439 ce &= ~(int64_t)Collation::CASE_MASK; // blank out case bits 440 int32_t index = binarySearch(uniqueCEs.getBuffer(), uniqueCEs.size(), ce); 441 U_ASSERT(index >= 0); 442 return miniCEs[index]; 443 } 444 445 UBool 446 CollationFastLatinBuilder::encodeUniqueCEs(UErrorCode &errorCode) { 447 if(U_FAILURE(errorCode)) { return FALSE; } 448 uprv_free(miniCEs); 449 miniCEs = (uint16_t *)uprv_malloc(uniqueCEs.size() * 2); 450 if(miniCEs == NULL) { 451 errorCode = U_MEMORY_ALLOCATION_ERROR; 452 return FALSE; 453 } 454 int32_t group = 1; 455 uint32_t lastGroupByte = result[group]; 456 // The lowest unique CE must be at least a secondary CE. 457 U_ASSERT(((uint32_t)uniqueCEs.elementAti(0) >> 16) != 0); 458 uint32_t prevPrimary = 0; 459 uint32_t prevSecondary = 0; 460 uint32_t pri = 0; 461 uint32_t sec = 0; 462 uint32_t ter = CollationFastLatin::COMMON_TER; 463 for(int32_t i = 0; i < uniqueCEs.size(); ++i) { 464 int64_t ce = uniqueCEs.elementAti(i); 465 // Note: At least one of the p/s/t weights changes from one unique CE to the next. 466 // (uniqueCEs does not store case bits.) 467 uint32_t p = (uint32_t)(ce >> 32); 468 if(p != prevPrimary) { 469 uint32_t p1 = p >> 24; 470 while(p1 > lastGroupByte) { 471 U_ASSERT(pri <= CollationFastLatin::MAX_LONG); 472 // Add the last "long primary" in or before the group 473 // into the upper 9 bits of the group entry. 474 result.setCharAt(group, (UChar)((pri << 4) | lastGroupByte)); 475 if(++group < headerLength) { // group is 1-based 476 lastGroupByte = result[group]; 477 } else { 478 lastGroupByte = 0xff; 479 break; 480 } 481 } 482 if(p < firstShortPrimary) { 483 if(pri == 0) { 484 pri = CollationFastLatin::MIN_LONG; 485 } else if(pri < CollationFastLatin::MAX_LONG) { 486 pri += CollationFastLatin::LONG_INC; 487 } else { 488 #if DEBUG_COLLATION_FAST_LATIN_BUILDER 489 printf("long-primary overflow for %08x\n", p); 490 #endif 491 miniCEs[i] = CollationFastLatin::BAIL_OUT; 492 continue; 493 } 494 } else { 495 if(pri < CollationFastLatin::MIN_SHORT) { 496 pri = CollationFastLatin::MIN_SHORT; 497 } else if(pri < (CollationFastLatin::MAX_SHORT - CollationFastLatin::SHORT_INC)) { 498 // Reserve the highest primary weight for U+FFFF. 499 pri += CollationFastLatin::SHORT_INC; 500 } else { 501 #if DEBUG_COLLATION_FAST_LATIN_BUILDER 502 printf("short-primary overflow for %08x\n", p); 503 #endif 504 shortPrimaryOverflow = TRUE; 505 miniCEs[i] = CollationFastLatin::BAIL_OUT; 506 continue; 507 } 508 } 509 prevPrimary = p; 510 prevSecondary = Collation::COMMON_WEIGHT16; 511 sec = CollationFastLatin::COMMON_SEC; 512 ter = CollationFastLatin::COMMON_TER; 513 } 514 uint32_t lower32 = (uint32_t)ce; 515 uint32_t s = lower32 >> 16; 516 if(s != prevSecondary) { 517 if(pri == 0) { 518 if(sec == 0) { 519 sec = CollationFastLatin::MIN_SEC_HIGH; 520 } else if(sec < CollationFastLatin::MAX_SEC_HIGH) { 521 sec += CollationFastLatin::SEC_INC; 522 } else { 523 miniCEs[i] = CollationFastLatin::BAIL_OUT; 524 continue; 525 } 526 prevSecondary = s; 527 ter = CollationFastLatin::COMMON_TER; 528 } else if(s < Collation::COMMON_WEIGHT16) { 529 if(sec == CollationFastLatin::COMMON_SEC) { 530 sec = CollationFastLatin::MIN_SEC_BEFORE; 531 } else if(sec < CollationFastLatin::MAX_SEC_BEFORE) { 532 sec += CollationFastLatin::SEC_INC; 533 } else { 534 miniCEs[i] = CollationFastLatin::BAIL_OUT; 535 continue; 536 } 537 } else if(s == Collation::COMMON_WEIGHT16) { 538 sec = CollationFastLatin::COMMON_SEC; 539 } else { 540 if(sec < CollationFastLatin::MIN_SEC_AFTER) { 541 sec = CollationFastLatin::MIN_SEC_AFTER; 542 } else if(sec < CollationFastLatin::MAX_SEC_AFTER) { 543 sec += CollationFastLatin::SEC_INC; 544 } else { 545 miniCEs[i] = CollationFastLatin::BAIL_OUT; 546 continue; 547 } 548 } 549 prevSecondary = s; 550 ter = CollationFastLatin::COMMON_TER; 551 } 552 U_ASSERT((lower32 & Collation::CASE_MASK) == 0); // blanked out in uniqueCEs 553 uint32_t t = lower32 & Collation::ONLY_TERTIARY_MASK; 554 if(t > Collation::COMMON_WEIGHT16) { 555 if(ter < CollationFastLatin::MAX_TER_AFTER) { 556 ++ter; 557 } else { 558 miniCEs[i] = CollationFastLatin::BAIL_OUT; 559 continue; 560 } 561 } 562 if(CollationFastLatin::MIN_LONG <= pri && pri <= CollationFastLatin::MAX_LONG) { 563 U_ASSERT(sec == CollationFastLatin::COMMON_SEC); 564 miniCEs[i] = (uint16_t)(pri | ter); 565 } else { 566 miniCEs[i] = (uint16_t)(pri | sec | ter); 567 } 568 } 569 #if DEBUG_COLLATION_FAST_LATIN_BUILDER 570 printf("last mini primary: %04x\n", pri); 571 #endif 572 #if DEBUG_COLLATION_FAST_LATIN_BUILDER >= 2 573 for(int32_t i = 0; i < uniqueCEs.size(); ++i) { 574 int64_t ce = uniqueCEs.elementAti(i); 575 printf("unique CE 0x%016lx -> 0x%04x\n", ce, miniCEs[i]); 576 } 577 #endif 578 return U_SUCCESS(errorCode); 579 } 580 581 UBool 582 CollationFastLatinBuilder::encodeCharCEs(UErrorCode &errorCode) { 583 if(U_FAILURE(errorCode)) { return FALSE; } 584 int32_t miniCEsStart = result.length(); 585 for(int32_t i = 0; i < CollationFastLatin::NUM_FAST_CHARS; ++i) { 586 result.append(0); // initialize to completely ignorable 587 } 588 int32_t indexBase = result.length(); 589 for(int32_t i = 0; i < CollationFastLatin::NUM_FAST_CHARS; ++i) { 590 int64_t ce = charCEs[i][0]; 591 if(isContractionCharCE(ce)) { continue; } // defer contraction 592 uint32_t miniCE = encodeTwoCEs(ce, charCEs[i][1]); 593 if(miniCE > 0xffff) { 594 // Note: There is a chance that this new expansion is the same as a previous one, 595 // and if so, then we could reuse the other expansion. 596 // However, that seems unlikely. 597 int32_t expansionIndex = result.length() - indexBase; 598 if(expansionIndex > (int32_t)CollationFastLatin::INDEX_MASK) { 599 miniCE = CollationFastLatin::BAIL_OUT; 600 } else { 601 result.append((UChar)(miniCE >> 16)).append((UChar)miniCE); 602 miniCE = CollationFastLatin::EXPANSION | expansionIndex; 603 } 604 } 605 result.setCharAt(miniCEsStart + i, (UChar)miniCE); 606 } 607 return U_SUCCESS(errorCode); 608 } 609 610 UBool 611 CollationFastLatinBuilder::encodeContractions(UErrorCode &errorCode) { 612 // We encode all contraction lists so that the first word of a list 613 // terminates the previous list, and we only need one additional terminator at the end. 614 if(U_FAILURE(errorCode)) { return FALSE; } 615 int32_t indexBase = headerLength + CollationFastLatin::NUM_FAST_CHARS; 616 int32_t firstContractionIndex = result.length(); 617 for(int32_t i = 0; i < CollationFastLatin::NUM_FAST_CHARS; ++i) { 618 int64_t ce = charCEs[i][0]; 619 if(!isContractionCharCE(ce)) { continue; } 620 int32_t contractionIndex = result.length() - indexBase; 621 if(contractionIndex > (int32_t)CollationFastLatin::INDEX_MASK) { 622 result.setCharAt(headerLength + i, CollationFastLatin::BAIL_OUT); 623 continue; 624 } 625 UBool firstTriple = TRUE; 626 for(int32_t index = (int32_t)ce & 0x7fffffff;; index += 3) { 627 int32_t x = contractionCEs.elementAti(index); 628 if((uint32_t)x == CollationFastLatin::CONTR_CHAR_MASK && !firstTriple) { break; } 629 int64_t cce0 = contractionCEs.elementAti(index + 1); 630 int64_t cce1 = contractionCEs.elementAti(index + 2); 631 uint32_t miniCE = encodeTwoCEs(cce0, cce1); 632 if(miniCE == CollationFastLatin::BAIL_OUT) { 633 result.append((UChar)(x | (1 << CollationFastLatin::CONTR_LENGTH_SHIFT))); 634 } else if(miniCE <= 0xffff) { 635 result.append((UChar)(x | (2 << CollationFastLatin::CONTR_LENGTH_SHIFT))); 636 result.append((UChar)miniCE); 637 } else { 638 result.append((UChar)(x | (3 << CollationFastLatin::CONTR_LENGTH_SHIFT))); 639 result.append((UChar)(miniCE >> 16)).append((UChar)miniCE); 640 } 641 firstTriple = FALSE; 642 } 643 // Note: There is a chance that this new contraction list is the same as a previous one, 644 // and if so, then we could truncate the result and reuse the other list. 645 // However, that seems unlikely. 646 result.setCharAt(headerLength + i, 647 (UChar)(CollationFastLatin::CONTRACTION | contractionIndex)); 648 } 649 if(result.length() > firstContractionIndex) { 650 // Terminate the last contraction list. 651 result.append((UChar)CollationFastLatin::CONTR_CHAR_MASK); 652 } 653 if(result.isBogus()) { 654 errorCode = U_MEMORY_ALLOCATION_ERROR; 655 return FALSE; 656 } 657 #if DEBUG_COLLATION_FAST_LATIN_BUILDER 658 printf("** fast Latin %d * 2 = %d bytes\n", result.length(), result.length() * 2); 659 puts(" header & below-digit groups map"); 660 int32_t i = 0; 661 for(; i < headerLength; ++i) { 662 printf(" %04x", result[i]); 663 } 664 printf("\n char mini CEs"); 665 U_ASSERT(CollationFastLatin::NUM_FAST_CHARS % 16 == 0); 666 for(; i < indexBase; i += 16) { 667 UChar32 c = i - headerLength; 668 if(c >= CollationFastLatin::LATIN_LIMIT) { 669 c = CollationFastLatin::PUNCT_START + c - CollationFastLatin::LATIN_LIMIT; 670 } 671 printf("\n %04x:", c); 672 for(int32_t j = 0; j < 16; ++j) { 673 printf(" %04x", result[i + j]); 674 } 675 } 676 printf("\n expansions & contractions"); 677 for(; i < result.length(); ++i) { 678 if((i - indexBase) % 16 == 0) { puts(""); } 679 printf(" %04x", result[i]); 680 } 681 puts(""); 682 #endif 683 return TRUE; 684 } 685 686 uint32_t 687 CollationFastLatinBuilder::encodeTwoCEs(int64_t first, int64_t second) const { 688 if(first == 0) { 689 return 0; // completely ignorable 690 } 691 if(first == Collation::NO_CE) { 692 return CollationFastLatin::BAIL_OUT; 693 } 694 U_ASSERT((uint32_t)(first >> 32) != Collation::NO_CE_PRIMARY); 695 696 uint32_t miniCE = getMiniCE(first); 697 if(miniCE == CollationFastLatin::BAIL_OUT) { return miniCE; } 698 if(miniCE >= CollationFastLatin::MIN_SHORT) { 699 // Extract & copy the case bits. 700 // Shift them from normal CE bits 15..14 to mini CE bits 4..3. 701 uint32_t c = (((uint32_t)first & Collation::CASE_MASK) >> (14 - 3)); 702 // Only in mini CEs: Ignorable case bits = 0, lowercase = 1. 703 c += CollationFastLatin::LOWER_CASE; 704 miniCE |= c; 705 } 706 if(second == 0) { return miniCE; } 707 708 uint32_t miniCE1 = getMiniCE(second); 709 if(miniCE1 == CollationFastLatin::BAIL_OUT) { return miniCE1; } 710 711 uint32_t case1 = (uint32_t)second & Collation::CASE_MASK; 712 if(miniCE >= CollationFastLatin::MIN_SHORT && 713 (miniCE & CollationFastLatin::SECONDARY_MASK) == CollationFastLatin::COMMON_SEC) { 714 // Try to combine the two mini CEs into one. 715 uint32_t sec1 = miniCE1 & CollationFastLatin::SECONDARY_MASK; 716 uint32_t ter1 = miniCE1 & CollationFastLatin::TERTIARY_MASK; 717 if(sec1 >= CollationFastLatin::MIN_SEC_HIGH && case1 == 0 && 718 ter1 == CollationFastLatin::COMMON_TER) { 719 // sec1>=sec_high implies pri1==0. 720 return (miniCE & ~CollationFastLatin::SECONDARY_MASK) | sec1; 721 } 722 } 723 724 if(miniCE1 <= CollationFastLatin::SECONDARY_MASK || CollationFastLatin::MIN_SHORT <= miniCE1) { 725 // Secondary CE, or a CE with a short primary, copy the case bits. 726 case1 = (case1 >> (14 - 3)) + CollationFastLatin::LOWER_CASE; 727 miniCE1 |= case1; 728 } 729 return (miniCE << 16) | miniCE1; 730 } 731 732 U_NAMESPACE_END 733 734 #endif // !UCONFIG_NO_COLLATION 735