Line data Source code
1 : // © 2016 and later: Unicode, Inc. and others.
2 : // License & terms of use: http://www.unicode.org/copyright.html
3 : /*
4 : *******************************************************************************
5 : * Copyright (C) 2013-2015, International Business Machines
6 : * Corporation and others. All Rights Reserved.
7 : *******************************************************************************
8 : * collationfastlatin.h
9 : *
10 : * created on: 2013aug09
11 : * created by: Markus W. Scherer
12 : */
13 :
14 : #ifndef __COLLATIONFASTLATIN_H__
15 : #define __COLLATIONFASTLATIN_H__
16 :
17 : #include "unicode/utypes.h"
18 :
19 : #if !UCONFIG_NO_COLLATION
20 :
21 : U_NAMESPACE_BEGIN
22 :
23 : struct CollationData;
24 : struct CollationSettings;
25 :
26 : class U_I18N_API CollationFastLatin /* all static */ {
27 : public:
28 : /**
29 : * Fast Latin format version (one byte 1..FF).
30 : * Must be incremented for any runtime-incompatible changes,
31 : * in particular, for changes to any of the following constants.
32 : *
33 : * When the major version number of the main data format changes,
34 : * we can reset this fast Latin version to 1.
35 : */
36 : static const uint16_t VERSION = 2;
37 :
38 : static const int32_t LATIN_MAX = 0x17f;
39 : static const int32_t LATIN_LIMIT = LATIN_MAX + 1;
40 :
41 : static const int32_t LATIN_MAX_UTF8_LEAD = 0xc5; // UTF-8 lead byte of LATIN_MAX
42 :
43 : static const int32_t PUNCT_START = 0x2000;
44 : static const int32_t PUNCT_LIMIT = 0x2040;
45 :
46 : // excludes U+FFFE & U+FFFF
47 : static const int32_t NUM_FAST_CHARS = LATIN_LIMIT + (PUNCT_LIMIT - PUNCT_START);
48 :
49 : // Note on the supported weight ranges:
50 : // Analysis of UCA 6.3 and CLDR 23 non-search tailorings shows that
51 : // the CEs for characters in the above ranges, excluding expansions with length >2,
52 : // excluding contractions of >2 characters, and other restrictions
53 : // (see the builder's getCEsFromCE32()),
54 : // use at most about 150 primary weights,
55 : // where about 94 primary weights are possibly-variable (space/punct/symbol/currency),
56 : // at most 4 secondary before-common weights,
57 : // at most 4 secondary after-common weights,
58 : // at most 16 secondary high weights (in secondary CEs), and
59 : // at most 4 tertiary after-common weights.
60 : // The following ranges are designed to support slightly more weights than that.
61 : // (en_US_POSIX is unusual: It creates about 64 variable + 116 Latin primaries.)
62 :
63 : // Digits may use long primaries (preserving more short ones)
64 : // or short primaries (faster) without changing this data structure.
65 : // (If we supported numeric collation, then digits would have to have long primaries
66 : // so that special handling does not affect the fast path.)
67 :
68 : static const uint32_t SHORT_PRIMARY_MASK = 0xfc00; // bits 15..10
69 : static const uint32_t INDEX_MASK = 0x3ff; // bits 9..0 for expansions & contractions
70 : static const uint32_t SECONDARY_MASK = 0x3e0; // bits 9..5
71 : static const uint32_t CASE_MASK = 0x18; // bits 4..3
72 : static const uint32_t LONG_PRIMARY_MASK = 0xfff8; // bits 15..3
73 : static const uint32_t TERTIARY_MASK = 7; // bits 2..0
74 : static const uint32_t CASE_AND_TERTIARY_MASK = CASE_MASK | TERTIARY_MASK;
75 :
76 : static const uint32_t TWO_SHORT_PRIMARIES_MASK =
77 : (SHORT_PRIMARY_MASK << 16) | SHORT_PRIMARY_MASK; // 0xfc00fc00
78 : static const uint32_t TWO_LONG_PRIMARIES_MASK =
79 : (LONG_PRIMARY_MASK << 16) | LONG_PRIMARY_MASK; // 0xfff8fff8
80 : static const uint32_t TWO_SECONDARIES_MASK =
81 : (SECONDARY_MASK << 16) | SECONDARY_MASK; // 0x3e003e0
82 : static const uint32_t TWO_CASES_MASK =
83 : (CASE_MASK << 16) | CASE_MASK; // 0x180018
84 : static const uint32_t TWO_TERTIARIES_MASK =
85 : (TERTIARY_MASK << 16) | TERTIARY_MASK; // 0x70007
86 :
87 : /**
88 : * Contraction with one fast Latin character.
89 : * Use INDEX_MASK to find the start of the contraction list after the fixed table.
90 : * The first entry contains the default mapping.
91 : * Otherwise use CONTR_CHAR_MASK for the contraction character index
92 : * (in ascending order).
93 : * Use CONTR_LENGTH_SHIFT for the length of the entry
94 : * (1=BAIL_OUT, 2=one CE, 3=two CEs).
95 : *
96 : * Also, U+0000 maps to a contraction entry, so that the fast path need not
97 : * check for NUL termination.
98 : * It usually maps to a contraction list with only the completely ignorable default value.
99 : */
100 : static const uint32_t CONTRACTION = 0x400;
101 : /**
102 : * An expansion encodes two CEs.
103 : * Use INDEX_MASK to find the pair of CEs after the fixed table.
104 : *
105 : * The higher a mini CE value, the easier it is to process.
106 : * For expansions and higher, no context needs to be considered.
107 : */
108 : static const uint32_t EXPANSION = 0x800;
109 : /**
110 : * Encodes one CE with a long/low mini primary (there are 128).
111 : * All potentially-variable primaries must be in this range,
112 : * to make the short-primary path as fast as possible.
113 : */
114 : static const uint32_t MIN_LONG = 0xc00;
115 : static const uint32_t LONG_INC = 8;
116 : static const uint32_t MAX_LONG = 0xff8;
117 : /**
118 : * Encodes one CE with a short/high primary (there are 60),
119 : * plus a secondary CE if the secondary weight is high.
120 : * Fast handling: At least all letter primaries should be in this range.
121 : */
122 : static const uint32_t MIN_SHORT = 0x1000;
123 : static const uint32_t SHORT_INC = 0x400;
124 : /** The highest primary weight is reserved for U+FFFF. */
125 : static const uint32_t MAX_SHORT = SHORT_PRIMARY_MASK;
126 :
127 : static const uint32_t MIN_SEC_BEFORE = 0; // must add SEC_OFFSET
128 : static const uint32_t SEC_INC = 0x20;
129 : static const uint32_t MAX_SEC_BEFORE = MIN_SEC_BEFORE + 4 * SEC_INC; // 5 before common
130 : static const uint32_t COMMON_SEC = MAX_SEC_BEFORE + SEC_INC;
131 : static const uint32_t MIN_SEC_AFTER = COMMON_SEC + SEC_INC;
132 : static const uint32_t MAX_SEC_AFTER = MIN_SEC_AFTER + 5 * SEC_INC; // 6 after common
133 : static const uint32_t MIN_SEC_HIGH = MAX_SEC_AFTER + SEC_INC; // 20 high secondaries
134 : static const uint32_t MAX_SEC_HIGH = SECONDARY_MASK;
135 :
136 : /**
137 : * Lookup: Add this offset to secondary weights, except for completely ignorable CEs.
138 : * Must be greater than any special value, e.g., MERGE_WEIGHT.
139 : * The exact value is not relevant for the format version.
140 : */
141 : static const uint32_t SEC_OFFSET = SEC_INC;
142 : static const uint32_t COMMON_SEC_PLUS_OFFSET = COMMON_SEC + SEC_OFFSET;
143 :
144 : static const uint32_t TWO_SEC_OFFSETS =
145 : (SEC_OFFSET << 16) | SEC_OFFSET; // 0x200020
146 : static const uint32_t TWO_COMMON_SEC_PLUS_OFFSET =
147 : (COMMON_SEC_PLUS_OFFSET << 16) | COMMON_SEC_PLUS_OFFSET;
148 :
149 : static const uint32_t LOWER_CASE = 8; // case bits include this offset
150 : static const uint32_t TWO_LOWER_CASES = (LOWER_CASE << 16) | LOWER_CASE; // 0x80008
151 :
152 : static const uint32_t COMMON_TER = 0; // must add TER_OFFSET
153 : static const uint32_t MAX_TER_AFTER = 7; // 7 after common
154 :
155 : /**
156 : * Lookup: Add this offset to tertiary weights, except for completely ignorable CEs.
157 : * Must be greater than any special value, e.g., MERGE_WEIGHT.
158 : * Must be greater than case bits as well, so that with combined case+tertiary weights
159 : * plus the offset the tertiary bits does not spill over into the case bits.
160 : * The exact value is not relevant for the format version.
161 : */
162 : static const uint32_t TER_OFFSET = SEC_OFFSET;
163 : static const uint32_t COMMON_TER_PLUS_OFFSET = COMMON_TER + TER_OFFSET;
164 :
165 : static const uint32_t TWO_TER_OFFSETS = (TER_OFFSET << 16) | TER_OFFSET;
166 : static const uint32_t TWO_COMMON_TER_PLUS_OFFSET =
167 : (COMMON_TER_PLUS_OFFSET << 16) | COMMON_TER_PLUS_OFFSET;
168 :
169 : static const uint32_t MERGE_WEIGHT = 3;
170 : static const uint32_t EOS = 2; // end of string
171 : static const uint32_t BAIL_OUT = 1;
172 :
173 : /**
174 : * Contraction result first word bits 8..0 contain the
175 : * second contraction character, as a char index 0..NUM_FAST_CHARS-1.
176 : * Each contraction list is terminated with a word containing CONTR_CHAR_MASK.
177 : */
178 : static const uint32_t CONTR_CHAR_MASK = 0x1ff;
179 : /**
180 : * Contraction result first word bits 10..9 contain the result length:
181 : * 1=bail out, 2=one mini CE, 3=two mini CEs
182 : */
183 : static const uint32_t CONTR_LENGTH_SHIFT = 9;
184 :
185 : /**
186 : * Comparison return value when the regular comparison must be used.
187 : * The exact value is not relevant for the format version.
188 : */
189 : static const int32_t BAIL_OUT_RESULT = -2;
190 :
191 0 : static inline int32_t getCharIndex(UChar c) {
192 0 : if(c <= LATIN_MAX) {
193 0 : return c;
194 0 : } else if(PUNCT_START <= c && c < PUNCT_LIMIT) {
195 0 : return c - (PUNCT_START - LATIN_LIMIT);
196 : } else {
197 : // Not a fast Latin character.
198 : // Note: U+FFFE & U+FFFF are forbidden in tailorings
199 : // and thus do not occur in any contractions.
200 0 : return -1;
201 : }
202 : }
203 :
204 : /**
205 : * Computes the options value for the compare functions
206 : * and writes the precomputed primary weights.
207 : * Returns -1 if the Latin fastpath is not supported for the data and settings.
208 : * The capacity must be LATIN_LIMIT.
209 : */
210 : static int32_t getOptions(const CollationData *data, const CollationSettings &settings,
211 : uint16_t *primaries, int32_t capacity);
212 :
213 : static int32_t compareUTF16(const uint16_t *table, const uint16_t *primaries, int32_t options,
214 : const UChar *left, int32_t leftLength,
215 : const UChar *right, int32_t rightLength);
216 :
217 : static int32_t compareUTF8(const uint16_t *table, const uint16_t *primaries, int32_t options,
218 : const uint8_t *left, int32_t leftLength,
219 : const uint8_t *right, int32_t rightLength);
220 :
221 : private:
222 : static uint32_t lookup(const uint16_t *table, UChar32 c);
223 : static uint32_t lookupUTF8(const uint16_t *table, UChar32 c,
224 : const uint8_t *s8, int32_t &sIndex, int32_t sLength);
225 : static uint32_t lookupUTF8Unsafe(const uint16_t *table, UChar32 c,
226 : const uint8_t *s8, int32_t &sIndex);
227 :
228 : static uint32_t nextPair(const uint16_t *table, UChar32 c, uint32_t ce,
229 : const UChar *s16, const uint8_t *s8, int32_t &sIndex, int32_t &sLength);
230 :
231 0 : static inline uint32_t getPrimaries(uint32_t variableTop, uint32_t pair) {
232 0 : uint32_t ce = pair & 0xffff;
233 0 : if(ce >= MIN_SHORT) { return pair & TWO_SHORT_PRIMARIES_MASK; }
234 0 : if(ce > variableTop) { return pair & TWO_LONG_PRIMARIES_MASK; }
235 0 : if(ce >= MIN_LONG) { return 0; } // variable
236 0 : return pair; // special mini CE
237 : }
238 0 : static inline uint32_t getSecondariesFromOneShortCE(uint32_t ce) {
239 0 : ce &= SECONDARY_MASK;
240 0 : if(ce < MIN_SEC_HIGH) {
241 0 : return ce + SEC_OFFSET;
242 : } else {
243 0 : return ((ce + SEC_OFFSET) << 16) | COMMON_SEC_PLUS_OFFSET;
244 : }
245 : }
246 : static uint32_t getSecondaries(uint32_t variableTop, uint32_t pair);
247 : static uint32_t getCases(uint32_t variableTop, UBool strengthIsPrimary, uint32_t pair);
248 : static uint32_t getTertiaries(uint32_t variableTop, UBool withCaseBits, uint32_t pair);
249 : static uint32_t getQuaternaries(uint32_t variableTop, uint32_t pair);
250 :
251 : private:
252 : CollationFastLatin(); // no constructor
253 : };
254 :
255 : /*
256 : * Format of the CollationFastLatin data table.
257 : * CollationFastLatin::VERSION = 2.
258 : *
259 : * This table contains data for a Latin-text collation fastpath.
260 : * The data is stored as an array of uint16_t which contains the following parts.
261 : *
262 : * uint16_t -- version & header length
263 : * Bits 15..8: version, must match the VERSION
264 : * 7..0: length of the header
265 : *
266 : * uint16_t varTops[header length - 1]
267 : * Version 2:
268 : * varTops[m] is the highest CollationFastLatin long-primary weight
269 : * of supported maxVariable group m
270 : * (special reorder group space, punct, symbol, currency).
271 : *
272 : * Version 1:
273 : * Each of these values maps the variable top lead byte of a supported maxVariable group
274 : * to the highest CollationFastLatin long-primary weight.
275 : * The values are stored in ascending order.
276 : * Bits 15..7: max fast-Latin long-primary weight (bits 11..3 shifted left by 4 bits)
277 : * 6..0: regular primary lead byte
278 : *
279 : * uint16_t miniCEs[0x1c0]
280 : * A mini collation element for each character U+0000..U+017F and U+2000..U+203F.
281 : * Each value encodes one or two mini CEs (two are possible if the first one
282 : * has a short mini primary and the second one is a secondary CE, i.e., primary == 0),
283 : * or points to an expansion or to a contraction table.
284 : * U+0000 always has a contraction entry,
285 : * so that NUL-termination need not be tested in the fastpath.
286 : * If the collation elements for a character or contraction cannot be encoded in this format,
287 : * then the BAIL_OUT value is stored.
288 : * For details see the comments for the class constants.
289 : *
290 : * uint16_t expansions[variable length];
291 : * Expansion mini CEs contain an offset relative to just after the miniCEs table.
292 : * An expansions contains exactly 2 mini CEs.
293 : *
294 : * uint16_t contractions[variable length];
295 : * Contraction mini CEs contain an offset relative to just after the miniCEs table.
296 : * It points to a list of tuples which map from a contraction suffix character to a result.
297 : * First uint16_t of each tuple:
298 : * Bits 10..9: Length of the result (1..3), see comments on CONTR_LENGTH_SHIFT.
299 : * Bits 8..0: Contraction character, see comments on CONTR_CHAR_MASK.
300 : * This is followed by 0, 1, or 2 uint16_t according to the length.
301 : * Each list is terminated by an entry with CONTR_CHAR_MASK.
302 : * Each list starts with such an entry which also contains the default result
303 : * for when there is no contraction match.
304 : *
305 : * -----------------
306 : * Changes for version 2 (ICU 55)
307 : *
308 : * Special reorder groups do not necessarily start on whole primary lead bytes any more.
309 : * Therefore, the varTops data has a new format:
310 : * Version 1 stored the lead bytes of the highest root primaries for
311 : * the maxVariable-supported special reorder groups.
312 : * Now the top 16 bits would need to be stored,
313 : * and it is simpler to store only the fast-Latin weights.
314 : */
315 :
316 : U_NAMESPACE_END
317 :
318 : #endif // !UCONFIG_NO_COLLATION
319 : #endif // __COLLATIONFASTLATIN_H__
|