Debian Codegraph @

src/backend/utils/adt/levenshtein.c

raw
1 /*-------------------------------------------------------------------------
2 *
3 * levenshtein.c
4 * Levenshtein distance implementation.
5 *
6 * Original author: Joe Conway <[email protected]>
7 *
8 * This file is included by varlena.c twice, to provide matching code for (1)
9 * Levenshtein distance with custom costings, and (2) Levenshtein distance with
10 * custom costings and a "max" value above which exact distances are not
11 * interesting. Before the inclusion, we rely on the presence of the inline
12 * functions rest_of_char_same() and levenshtein_result().
13 *
14 * Written based on a description of the algorithm by Michael Gilleland found
15 * at http://www.merriampark.com/ld.htm. Also looked at levenshtein.c in the
16 * PHP 4.0.6 distribution for inspiration. Configurable penalty costs
17 * extension is introduced by Volkan YAZICI <[email protected].
18 *
19 * Copyright (c) 2001-2025, PostgreSQL Global Development Group
20 *
21 * IDENTIFICATION
22 * src/backend/utils/adt/levenshtein.c
23 *
24 *-------------------------------------------------------------------------
25 */
26 #define MAX_LEVENSHTEIN_STRLEN 255
27
28 /*
29 * Calculates Levenshtein distance metric between supplied strings, which are
30 * not necessarily null-terminated.
31 *
32 * source: source string, of length slen bytes.
33 * target: target string, of length tlen bytes.
34 * ins_c, del_c, sub_c: costs to charge for character insertion, deletion,
35 * and substitution respectively; (1, 1, 1) costs suffice for common
36 * cases, but your mileage may vary.
37 * max_d: if provided and >= 0, maximum distance we care about; see below.
38 * trusted: caller is trusted and need not obey MAX_LEVENSHTEIN_STRLEN.
39 *
40 * One way to compute Levenshtein distance is to incrementally construct
41 * an (m+1)x(n+1) matrix where cell (i, j) represents the minimum number
42 * of operations required to transform the first i characters of s into
43 * the first j characters of t. The last column of the final row is the
44 * answer.
45 *
46 * We use that algorithm here with some modification. In lieu of holding
47 * the entire array in memory at once, we'll just use two arrays of size
48 * m+1 for storing accumulated values. At each step one array represents
49 * the "previous" row and one is the "current" row of the notional large
50 * array.
51 *
52 * If max_d >= 0, we only need to provide an accurate answer when that answer
53 * is less than or equal to max_d. From any cell in the matrix, there is
54 * theoretical "minimum residual distance" from that cell to the last column
55 * of the final row. This minimum residual distance is zero when the
56 * untransformed portions of the strings are of equal length (because we might
57 * get lucky and find all the remaining characters matching) and is otherwise
58 * based on the minimum number of insertions or deletions needed to make them
59 * equal length. The residual distance grows as we move toward the upper
60 * right or lower left corners of the matrix. When the max_d bound is
61 * usefully tight, we can use this property to avoid computing the entirety
62 * of each row; instead, we maintain a start_column and stop_column that
63 * identify the portion of the matrix close to the diagonal which can still
64 * affect the final answer.
65 */
66 int
67 #ifdef LEVENSHTEIN_LESS_EQUAL
68 varstr_levenshtein_less_equal(const char *source, int slen,
69 const char *target, int tlen,
70 int ins_c, int del_c, int sub_c,
71 int max_d, bool trusted)
72 #else
73 varstr_levenshtein(const char *source, int slen,
74 const char *target, int tlen,
75 int ins_c, int del_c, int sub_c,
76 bool trusted)
77 #endif
78 {
79 int m,
80 n;
81 int64 *prev;
82 int64 *curr;
83 int *s_char_len = NULL;
84 int j;
85 const char *y;
86 const char *send = source + slen;
87 const char *tend = target + tlen;
88 int64 ins_c_64 = ins_c;
89 int64 del_c_64 = del_c;
90 int64 sub_c_64 = sub_c;
91
92 /*
93 * For varstr_levenshtein_less_equal, we have real variables called
94 * start_column and stop_column; otherwise it's just short-hand for 0 and
95 * m.
96 */
97 #ifdef LEVENSHTEIN_LESS_EQUAL
98 int start_column,
99 stop_column;
100
101 #undef START_COLUMN
102 #undef STOP_COLUMN
103 #define START_COLUMN start_column
104 #define STOP_COLUMN stop_column
105 #else
106 #undef START_COLUMN
107 #undef STOP_COLUMN
108 #define START_COLUMN 0
109 #define STOP_COLUMN m
110 #endif
111
112 /* Convert string lengths (in bytes) to lengths in characters */
113 m = pg_mbstrlen_with_len(source, slen);
114 n = pg_mbstrlen_with_len(target, tlen);
115
116 /*
117 * We can transform an empty s into t with n insertions, or a non-empty t
118 * into an empty s with m deletions.
119 */
120 if (!m)
121 return levenshtein_result(n * ins_c_64);
122 if (!n)
123 return levenshtein_result(m * del_c_64);
124
125 /*
126 * For security concerns, restrict excessive CPU+RAM usage. (This
127 * implementation uses O(m) memory and has O(mn) complexity.) If
128 * "trusted" is true, caller is responsible for not making excessive
129 * requests, typically by using a small max_d along with strings that are
130 * bounded, though not necessarily to MAX_LEVENSHTEIN_STRLEN exactly.
131 */
132 if (!trusted &&
133 (m > MAX_LEVENSHTEIN_STRLEN ||
134 n > MAX_LEVENSHTEIN_STRLEN))
135 ereport(ERROR,
136 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
137 errmsg("levenshtein argument exceeds maximum length of %d characters",
138 MAX_LEVENSHTEIN_STRLEN)));
139
140 #ifdef LEVENSHTEIN_LESS_EQUAL
141 /* Initialize start and stop columns. */
142 start_column = 0;
143 stop_column = m + 1;
144
145 /*
146 * If max_d >= 0, determine whether the bound is impossibly tight. If so,
147 * return max_d + 1 immediately. Otherwise, determine whether it's tight
148 * enough to limit the computation we must perform. If so, figure out
149 * initial stop column.
150 */
151 if (max_d >= 0)
152 {
153 int64 min_theo_d; /* Theoretical minimum distance. */
154 int64 max_theo_d; /* Theoretical maximum distance. */
155 int net_inserts = n - m;
156
157 min_theo_d = net_inserts < 0 ?
158 -net_inserts * del_c_64 : net_inserts * ins_c_64;
159 if (min_theo_d > max_d)
160 return levenshtein_result((int64) max_d + 1);
161 if (ins_c_64 + del_c_64 < sub_c_64)
162 sub_c_64 = ins_c_64 + del_c_64;
163 max_theo_d = min_theo_d + sub_c_64 * Min(m, n);
164 if (max_d >= max_theo_d)
165 max_d = -1;
166 else if (ins_c_64 + del_c_64 > 0)
167 {
168 /*
169 * Figure out how much of the first row of the notional matrix we
170 * need to fill in. If the string is growing, the theoretical
171 * minimum distance already incorporates the cost of deleting the
172 * number of characters necessary to make the two strings equal in
173 * length. Each additional deletion forces another insertion, so
174 * the best-case total cost increases by ins_c + del_c. If the
175 * string is shrinking, the minimum theoretical cost assumes no
176 * excess deletions; that is, we're starting no further right than
177 * column n - m. If we do start further right, the best-case
178 * total cost increases by ins_c + del_c for each move right.
179 */
180 int64 slack_d = max_d - min_theo_d;
181 int best_column = net_inserts < 0 ? -net_inserts : 0;
182 int64 tmp;
183
184 tmp = best_column + (slack_d / (ins_c_64 + del_c_64)) + 1;
185 stop_column = Min(tmp, m + 1);
186 }
187 }
188 #endif
189
190 /*
191 * In order to avoid calling pg_mblen_range() repeatedly on each character
192 * in s, we cache all the lengths before starting the main loop -- but if
193 * all the characters in both strings are single byte, then we skip this
194 * and use a fast-path in the main loop. If only one string contains
195 * multi-byte characters, we still build the array, so that the fast-path
196 * needn't deal with the case where the array hasn't been initialized.
197 */
198 if (m != slen || n != tlen)
199 {
200 int i;
201 const char *cp = source;
202
203 s_char_len = (int *) palloc((m + 1) * sizeof(int));
204 for (i = 0; i < m; ++i)
205 {
206 s_char_len[i] = pg_mblen_range(cp, send);
207 cp += s_char_len[i];
208 }
209 s_char_len[i] = 0;
210 }
211
212 /* One more cell for initialization column and row. */
213 ++m;
214 ++n;
215
216 /* Previous and current rows of notional array. */
217 prev = (int64 *) palloc(2 * m * sizeof(int64));
218 curr = prev + m;
219
220 /*
221 * To transform the first i characters of s into the first 0 characters of
222 * t, we must perform i deletions.
223 */
224 for (int i = START_COLUMN; i < STOP_COLUMN; i++)
225 prev[i] = i * del_c_64;
226
227 /* Loop through rows of the notional array */
228 for (y = target, j = 1; j < n; j++)
229 {
230 int64 *temp;
231 const char *x = source;
232 int y_char_len = n != tlen + 1 ? pg_mblen_range(y, tend) : 1;
233 int i;
234
235 #ifdef LEVENSHTEIN_LESS_EQUAL
236
237 /*
238 * In the best case, values percolate down the diagonal unchanged, so
239 * we must increment stop_column unless it's already on the right end
240 * of the array. The inner loop will read prev[stop_column], so we
241 * have to initialize it even though it shouldn't affect the result.
242 */
243 if (stop_column < m)
244 {
245 prev[stop_column] = (int64) max_d + 1;
246 ++stop_column;
247 }
248
249 /*
250 * The main loop fills in curr, but curr[0] needs a special case: to
251 * transform the first 0 characters of s into the first j characters
252 * of t, we must perform j insertions. However, if start_column > 0,
253 * this special case does not apply.
254 */
255 if (start_column == 0)
256 {
257 curr[0] = j * ins_c_64;
258 i = 1;
259 }
260 else
261 i = start_column;
262 #else
263 curr[0] = j * ins_c_64;
264 i = 1;
265 #endif
266
267 /*
268 * This inner loop is critical to performance, so we include a
269 * fast-path to handle the (fairly common) case where no multibyte
270 * characters are in the mix. The fast-path is entitled to assume
271 * that if s_char_len is not initialized then BOTH strings contain
272 * only single-byte characters.
273 */
274 if (s_char_len != NULL)
275 {
276 for (; i < STOP_COLUMN; i++)
277 {
278 int64 ins;
279 int64 del;
280 int64 sub;
281 int x_char_len = s_char_len[i - 1];
282
283 /*
284 * Calculate costs for insertion, deletion, and substitution.
285 *
286 * When calculating cost for substitution, we compare the last
287 * character of each possibly-multibyte character first,
288 * because that's enough to rule out most mis-matches. If we
289 * get past that test, then we compare the lengths and the
290 * remaining bytes.
291 */
292 ins = prev[i] + ins_c_64;
293 del = curr[i - 1] + del_c_64;
294 if (x[x_char_len - 1] == y[y_char_len - 1]
295 && x_char_len == y_char_len &&
296 (x_char_len == 1 || rest_of_char_same(x, y, x_char_len)))
297 sub = prev[i - 1];
298 else
299 sub = prev[i - 1] + sub_c_64;
300
301 /* Take the one with minimum cost. */
302 curr[i] = Min(ins, del);
303 curr[i] = Min(curr[i], sub);
304
305 /* Point to next character. */
306 x += x_char_len;
307 }
308 }
309 else
310 {
311 for (; i < STOP_COLUMN; i++)
312 {
313 int64 ins;
314 int64 del;
315 int64 sub;
316
317 /* Calculate costs for insertion, deletion, and substitution. */
318 ins = prev[i] + ins_c_64;
319 del = curr[i - 1] + del_c_64;
320 sub = prev[i - 1] + ((*x == *y) ? 0 : sub_c_64);
321
322 /* Take the one with minimum cost. */
323 curr[i] = Min(ins, del);
324 curr[i] = Min(curr[i], sub);
325
326 /* Point to next character. */
327 x++;
328 }
329 }
330
331 /* Swap current row with previous row. */
332 temp = curr;
333 curr = prev;
334 prev = temp;
335
336 /* Point to next character. */
337 y += y_char_len;
338
339 #ifdef LEVENSHTEIN_LESS_EQUAL
340
341 /*
342 * This chunk of code represents a significant performance hit if used
343 * in the case where there is no max_d bound. This is probably not
344 * because the max_d >= 0 test itself is expensive, but rather because
345 * the possibility of needing to execute this code prevents tight
346 * optimization of the loop as a whole.
347 */
348 if (max_d >= 0)
349 {
350 /*
351 * The "zero point" is the column of the current row where the
352 * remaining portions of the strings are of equal length. There
353 * are (n - 1) characters in the target string, of which j have
354 * been transformed. There are (m - 1) characters in the source
355 * string, so we want to find the value for zp where (n - 1) - j =
356 * (m - 1) - zp.
357 */
358 int zp = j - (n - m);
359
360 /* Check whether the stop column can slide left. */
361 while (stop_column > 0)
362 {
363 int ii = stop_column - 1;
364 int net_inserts = ii - zp;
365
366 if (prev[ii] + (net_inserts > 0 ? net_inserts * ins_c_64 :
367 -net_inserts * del_c_64) <= max_d)
368 break;
369 stop_column--;
370 }
371
372 /* Check whether the start column can slide right. */
373 while (start_column < stop_column)
374 {
375 int net_inserts = start_column - zp;
376
377 if (prev[start_column] +
378 (net_inserts > 0 ? net_inserts * ins_c_64 :
379 -net_inserts * del_c_64) <= max_d)
380 break;
381
382 /*
383 * We'll never again update these values, so we must make sure
384 * there's nothing here that could confuse any future
385 * iteration of the outer loop.
386 */
387 prev[start_column] = (int64) max_d + 1;
388 curr[start_column] = (int64) max_d + 1;
389 if (start_column != 0)
390 source += (s_char_len != NULL) ? s_char_len[start_column - 1] : 1;
391 start_column++;
392 }
393
394 /* If they cross, we're going to exceed the bound. */
395 if (start_column >= stop_column)
396 return levenshtein_result((int64) max_d + 1);
397 }
398 #endif
399 }
400
401 /*
402 * Because the final value was swapped from the previous row to the
403 * current row, that's where we'll find it.
404 */
405 return levenshtein_result(prev[m - 1]);
406 }