1 #include "parse_trigrams.h"
3 #include "unique_sort.h"
11 string print_td(const TrigramDisjunction &td)
13 if (td.read_trigrams.size() == 0) {
14 // Before we've done hash lookups (or none matched), so print all alternatives.
15 if (td.trigram_alternatives.size() == 1) {
16 return print_trigram(td.trigram_alternatives[0]);
21 for (uint32_t trgm : td.trigram_alternatives) {
24 ret += print_trigram(trgm);
30 // Print only those that we actually have in the index.
31 if (td.read_trigrams.size() == 1) {
32 return print_trigram(td.read_trigrams[0].first.trgm);
37 for (auto &[trgmptr, len] : td.read_trigrams) {
40 ret += print_trigram(trgmptr.trgm);
48 string print_trigram(uint32_t trgm)
51 char(trgm & 0xff), char((trgm >> 8) & 0xff), char((trgm >> 16) & 0xff)
55 for (unsigned i = 0; i < 3;) {
60 } else if (int(ch[i]) >= 32 && int(ch[i]) <= 127) { // Holds no matter whether char is signed or unsigned.
64 // See if we have an entire UTF-8 codepoint, and that it's reasonably printable.
65 mbtowc(nullptr, 0, 0);
67 int ret = mbtowc(&pwc, ch + i, 3 - i);
68 if (ret >= 1 && pwc >= 32) {
69 str.append(ch + i, ret);
73 snprintf(buf, sizeof(buf), "\\x{%02x}", (unsigned char)ch[i]);
83 pair<uint32_t, size_t> read_unigram(const string &s, size_t start)
85 if (start >= s.size()) {
86 return { PREMATURE_END_UNIGRAM, 0 };
88 if (s[start] == '\\') {
90 if (start + 1 >= s.size()) {
91 return { PREMATURE_END_UNIGRAM, 1 };
93 return { (unsigned char)s[start + 1], 2 };
96 if (s[start] == '*' || s[start] == '?') {
98 return { WILDCARD_UNIGRAM, 1 };
100 if (s[start] == '[') {
101 // Character class; search to find the end.
103 if (start + len >= s.size()) {
104 return { PREMATURE_END_UNIGRAM, len };
106 if (s[start + len] == '!') {
109 if (start + len >= s.size()) {
110 return { PREMATURE_END_UNIGRAM, len };
112 if (s[start + len] == ']') {
116 if (start + len >= s.size()) {
117 return { PREMATURE_END_UNIGRAM, len };
119 if (s[start + len] == ']') {
120 return { WILDCARD_UNIGRAM, len + 1 };
127 return { (unsigned char)s[start], 1 };
130 uint32_t read_trigram(const string &s, size_t start)
132 pair<uint32_t, size_t> u1 = read_unigram(s, start);
133 if (u1.first == WILDCARD_UNIGRAM || u1.first == PREMATURE_END_UNIGRAM) {
136 pair<uint32_t, size_t> u2 = read_unigram(s, start + u1.second);
137 if (u2.first == WILDCARD_UNIGRAM || u2.first == PREMATURE_END_UNIGRAM) {
140 pair<uint32_t, size_t> u3 = read_unigram(s, start + u1.second + u2.second);
141 if (u3.first == WILDCARD_UNIGRAM || u3.first == PREMATURE_END_UNIGRAM) {
144 return u1.first | (u2.first << 8) | (u3.first << 16);
147 struct TrigramState {
149 unsigned next_codepoint;
151 bool operator<(const TrigramState &other) const
153 if (next_codepoint != other.next_codepoint)
154 return next_codepoint < other.next_codepoint;
155 return buffered < other.buffered;
157 bool operator==(const TrigramState &other) const
159 return next_codepoint == other.next_codepoint &&
160 buffered == other.buffered;
164 void parse_trigrams_ignore_case(const string &needle, vector<TrigramDisjunction> *trigram_groups)
166 vector<vector<string>> alternatives_for_cp;
168 // Parse the needle into Unicode code points, and do inverse case folding
169 // on each to find legal alternatives. This is far from perfect (e.g. ß
170 // will not become ss), but it's generally the best we can do without
171 // involving ICU or the likes.
172 mbtowc(nullptr, 0, 0);
173 const char *ptr = needle.c_str();
174 while (*ptr != '\0') {
176 int ret = mbtowc(&ch, ptr, strlen(ptr));
182 char buf[MB_CUR_MAX];
184 alt.push_back(string(ptr, ret));
186 if (towlower(ch) != wint_t(ch)) {
187 ret = wctomb(buf, towlower(ch));
188 alt.push_back(string(buf, ret));
190 if (towupper(ch) != wint_t(ch) && towupper(ch) != towlower(ch)) {
191 ret = wctomb(buf, towupper(ch));
192 alt.push_back(string(buf, ret));
194 alternatives_for_cp.push_back(move(alt));
197 // Now generate all possible byte strings from those code points in order;
198 // e.g., from abc, we'd create a and A, then extend those to ab aB Ab AB,
199 // then abc abC aBc aBC and so on. Since we don't want to have 2^n
200 // (or even 3^n) strings, we only extend them far enough to cover at
201 // least three bytes; this will give us a set of candidate trigrams
202 // (the filename must have at least one of those), and then we can
203 // chop off the first byte, deduplicate states and continue extending
204 // and generating trigram sets.
206 // There are a few special cases, notably the dotted i (İ), where the
207 // UTF-8 versions of upper and lower case have different number of bytes.
208 // If this happens, we can have combinatorial explosion and get many more
209 // than the normal 8 states. We detect this and simply bomb out; it will
210 // never really happen in real strings, and stopping trigram generation
211 // really only means our pruning of candidates will be less effective.
212 vector<TrigramState> states;
213 states.push_back(TrigramState{ "", 0 });
216 // Extend every state so that it has buffered at least three bytes.
217 // If this isn't possible, we are done with the string (can generate
218 // no more trigrams).
219 bool need_another_pass;
221 need_another_pass = false;
222 vector<TrigramState> new_states;
223 for (const TrigramState &state : states) {
224 if (read_trigram(state.buffered, 0) != PREMATURE_END_UNIGRAM) {
225 // No need to extend this further.
226 new_states.push_back(state);
229 if (state.next_codepoint == alternatives_for_cp.size()) {
230 // We can't form a complete trigram from this alternative,
234 for (const string &rune : alternatives_for_cp[state.next_codepoint]) {
235 TrigramState new_state{ state.buffered + rune, state.next_codepoint + 1 };
236 if (read_trigram(state.buffered, 0) == PREMATURE_END_UNIGRAM) {
237 need_another_pass = true;
239 new_states.push_back(move(new_state));
242 states = move(new_states);
243 } while (need_another_pass);
245 // OK, so now we have a bunch of states, and all of them are at least
246 // three bytes long. This means we have a complete set of trigrams,
247 // and the destination filename must contain at least one of them.
248 // Output those trigrams, cut out the first byte and then deduplicate
249 // the states before we continue.
250 bool any_wildcard = false;
251 vector<uint32_t> trigram_alternatives;
252 for (TrigramState &state : states) {
253 trigram_alternatives.push_back(read_trigram(state.buffered, 0));
254 state.buffered.erase(0, read_unigram(state.buffered, 0).second);
255 assert(trigram_alternatives.back() != PREMATURE_END_UNIGRAM);
256 if (trigram_alternatives.back() == WILDCARD_UNIGRAM) {
257 // If any of the candidates are wildcards, we need to drop the entire OR group.
258 // (Most likely, all of them would be anyway.) We need to keep stripping out
259 // the first unigram from each state.
263 unique_sort(&trigram_alternatives); // Could have duplicates, although it's rare.
264 unique_sort(&states);
267 TrigramDisjunction new_pt;
268 new_pt.remaining_trigrams_to_read = trigram_alternatives.size();
269 new_pt.trigram_alternatives = move(trigram_alternatives);
270 new_pt.max_num_docids = 0;
271 trigram_groups->push_back(move(new_pt));
274 if (states.size() > 100) {
275 // A completely crazy pattern with lots of those special characters.
276 // We just give up; this isn't a realistic scenario anyway.
277 // We already have lots of trigrams that should reduce the amount of
284 void parse_trigrams(const string &needle, bool ignore_case, vector<TrigramDisjunction> *trigram_groups)
287 parse_trigrams_ignore_case(needle, trigram_groups);
291 // The case-sensitive case is straightforward.
292 for (size_t i = 0; i < needle.size(); i += read_unigram(needle, i).second) {
293 uint32_t trgm = read_trigram(needle, i);
294 if (trgm == WILDCARD_UNIGRAM || trgm == PREMATURE_END_UNIGRAM) {
295 // Invalid trigram, so skip.
299 TrigramDisjunction new_pt;
300 new_pt.remaining_trigrams_to_read = 1;
301 new_pt.trigram_alternatives.push_back(trgm);
302 new_pt.max_num_docids = 0;
303 trigram_groups->push_back(move(new_pt));