2 * Copyright (C) 2012, 2013
5 * Permission is hereby granted, free of charge, to any person obtaining a copy of
6 * this software and associated documentation files (the "Software"), to deal in
7 * the Software without restriction, including without limitation the rights to
8 * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
9 * of the Software, and to permit persons to whom the Software is furnished to do
10 * so, subject to the following conditions:
12 * The above copyright notice and this permission notice shall be included in all
13 * copies or substantial portions of the Software.
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
26 * This is a very clever method for correcting mistakes in QuakeC code
27 * most notably when invalid identifiers are used or inproper assignments;
28 * we can proprly lookup in multiple dictonaries (depening on the rules
29 * of what the task is trying to acomplish) to find the best possible
33 * A little about how it works, and probability theory:
35 * When given an identifier (which we will denote I), we're essentially
36 * just trying to choose the most likely correction for that identifier.
37 * (the actual "correction" can very well be the identifier itself).
38 * There is actually no way to know for sure that certian identifers
39 * such as "lates", need to be corrected to "late" or "latest" or any
40 * other permutations that look lexically the same. This is why we
41 * must advocate the usage of probabilities. This means that instead of
42 * just guessing, instead we're trying to find the correction for C,
43 * out of all possible corrections that maximizes the probability of C
44 * for the original identifer I.
46 * Bayes' Therom suggests something of the following:
47 * AC P(I|C) P(C) / P(I)
48 * Since P(I) is the same for every possibly I, we can ignore it giving
51 * This greatly helps visualize how the parts of the expression are performed
52 * there is essentially three, from right to left we perform the following:
54 * 1: P(C), the probability that a proposed correction C will stand on its
55 * own. This is called the language model.
57 * 2: P(I|C), the probability that I would be used, when the programmer
58 * really meant C. This is the error model.
60 * 3: AC, the control mechanisim, an enumerator if you will, one that
61 * enumerates all feasible values of C, to determine the one that
62 * gives the greatest probability score.
64 * In reality the requirement for a more complex expression involving
65 * two seperate models is considerably a waste. But one must recognize
66 * that P(C|I) is already conflating two factors. It's just much simpler
67 * to seperate the two models and deal with them explicitaly. To properly
68 * estimate P(C|I) you have to consider both the probability of C and
69 * probability of the transposition from C to I. It's simply much more
70 * cleaner, and direct to seperate the two factors.
72 * Research tells us that 80% to 95% of all spelling errors have an edit
73 * distance no greater than one. Knowing this we can optimize for most
74 * cases of mistakes without taking a performance hit. Which is what we
75 * base longer edit distances off of. Opposed to the original method of
76 * I had concieved of checking everything.
78 * A little information on additional algorithms used:
80 * Initially when I implemented this corrector, it was very slow.
81 * Need I remind you this is essentially a brute force attack on strings,
82 * and since every transformation requires dynamic memory allocations,
83 * you can easily imagine where most of the runtime conflated. Yes
84 * It went right to malloc. More than THREE MILLION malloc calls are
85 * performed for an identifier about 16 bytes long. This was such a
86 * shock to me. A forward allocator (or as some call it a bump-point
87 * allocator, or just a memory pool) was implemented. To combat this.
89 * But of course even other factors were making it slow. Initially
90 * this used a hashtable. And hashtables have a good constant lookup
91 * time complexity. But the problem wasn't in the hashtable, it was
92 * in the hashing (despite having one of the fastest hash functions
93 * known). Remember those 3 million mallocs? Well for every malloc
94 * there is also a hash. After 3 million hashes .. you start to get
95 * very slow. To combat this I had suggested burst tries to Blub.
96 * The next day he had implemented them. Sure enough this brought
97 * down the runtime by a factory > 100%
99 * Future Work (If we really need it)
101 * Currently we can only distinguishes one source of error in the
102 * language model we use. This could become an issue for identifiers
103 * that have close colliding rates, e.g colate->coat yields collate.
105 * Currently the error model has been fairly trivial, the smaller the
106 * edit distance the smaller the error. This usually causes some un-
107 * expected problems. e.g reciet->recite yields recipt. For QuakeC
108 * this could become a problem when lots of identifiers are involved.
110 * Our control mechanisim could use a limit, i.e limit the number of
111 * sets of edits for distance X. This would also increase execution
112 * speed considerably.
116 #define CORRECT_POOLSIZE (128*1024*1024)
118 * A forward allcator for the corrector. This corrector requires a lot
119 * of allocations. This forward allocator combats all those allocations
120 * and speeds us up a little. It also saves us space in a way since each
121 * allocation isn't wasting a little header space for when NOTRACK isn't
124 static unsigned char **correct_pool_data = NULL;
125 static unsigned char *correct_pool_this = NULL;
126 static size_t correct_pool_addr = 0;
128 static GMQCC_INLINE void correct_pool_new(void) {
129 correct_pool_addr = 0;
130 correct_pool_this = (unsigned char *)mem_a(CORRECT_POOLSIZE);
132 vec_push(correct_pool_data, correct_pool_this);
135 static GMQCC_INLINE void *correct_pool_alloc(size_t bytes) {
137 if (correct_pool_addr + bytes >= CORRECT_POOLSIZE)
140 data = correct_pool_this;
141 correct_pool_this += bytes;
142 correct_pool_addr += bytes;
147 static GMQCC_INLINE void correct_pool_delete(void) {
149 for (i = 0; i < vec_size(correct_pool_data); ++i)
150 mem_d(correct_pool_data[i]);
152 correct_pool_data = NULL;
153 correct_pool_this = NULL;
154 correct_pool_addr = 0;
158 static GMQCC_INLINE char *correct_pool_claim(const char *data) {
159 char *claim = util_strdup(data);
160 correct_pool_delete();
165 * A fast space efficent trie for a disctonary of identifiers. This is
166 * faster than a hashtable for one reason. A hashtable itself may have
167 * fast constant lookup time, but the hash itself must be very fast. We
168 * have one of the fastest hash functions for strings, but if you do a
169 * lost of hashing (which we do, almost 3 million hashes per identifier)
170 * a hashtable becomes slow. Very Very Slow.
172 correct_trie_t* correct_trie_new() {
173 correct_trie_t *t = (correct_trie_t*)mem_a(sizeof(correct_trie_t));
179 void correct_trie_del_sub(correct_trie_t *t) {
181 for (i = 0; i < vec_size(t->entries); ++i)
182 correct_trie_del_sub(&t->entries[i]);
183 vec_free(t->entries);
186 void correct_trie_del(correct_trie_t *t) {
188 for (i = 0; i < vec_size(t->entries); ++i)
189 correct_trie_del_sub(&t->entries[i]);
190 vec_free(t->entries);
194 void* correct_trie_get(const correct_trie_t *t, const char *key) {
195 const unsigned char *data = (const unsigned char*)key;
197 unsigned char ch = *data;
198 const size_t vs = vec_size(t->entries);
200 const correct_trie_t *entries = t->entries;
201 for (i = 0; i < vs; ++i) {
202 if (entries[i].ch == ch) {
214 void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
215 const unsigned char *data = (const unsigned char*)key;
217 const size_t vs = vec_size(t->entries);
218 unsigned char ch = *data;
219 correct_trie_t *entries = t->entries;
222 for (i = 0; i < vs; ++i) {
223 if (entries[i].ch == ch) {
229 correct_trie_t *elem = (correct_trie_t*)vec_add(t->entries, 1);
233 elem->entries = NULL;
243 * Implementation of the corrector algorithm commences. A very efficent
244 * brute-force attack (thanks to tries and mempool :-)).
246 static size_t *correct_find(correct_trie_t *table, const char *word) {
247 return (size_t*)correct_trie_get(table, word);
250 static int correct_update(correct_trie_t* *table, const char *word) {
251 size_t *data = correct_find(*table, word);
259 void correct_add(correct_trie_t* table, size_t ***size, const char *ident) {
261 const char *add = ident;
263 if (!correct_update(&table, add)) {
264 data = (size_t*)mem_a(sizeof(size_t));
267 vec_push((*size), data);
268 correct_trie_set(table, add, data);
272 void correct_del(correct_trie_t* dictonary, size_t **data) {
274 const size_t vs = vec_size(data);
276 for (i = 0; i < vs; i++)
280 correct_trie_del(dictonary);
284 * _ is valid in identifiers. I've yet to implement numerics however
285 * because they're only valid after the first character is of a _, or
288 static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyz"
289 "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
290 "_"; /* TODO: Numbers ... */
293 * correcting logic for the following forms of transformations:
299 static size_t correct_deletion(const char *ident, char **array, size_t index) {
301 const size_t len = strlen(ident);
303 for (; itr < len; itr++) {
304 char *a = (char*)correct_pool_alloc(len+1);
305 memcpy(a, ident, itr);
306 memcpy(a + itr, ident + itr + 1, len - itr);
307 array[index + itr] = a;
313 static size_t correct_transposition(const char *ident, char **array, size_t index) {
315 const size_t len = strlen(ident);
317 for (; itr < len - 1; itr++) {
319 char *a = (char*)correct_pool_alloc(len+1);
320 memcpy(a, ident, len+1);
324 array[index + itr] = a;
330 static size_t correct_alteration(const char *ident, char **array, size_t index) {
334 const size_t len = strlen(ident);
336 for (; itr < len; itr++) {
337 for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
338 char *a = (char*)correct_pool_alloc(len+1);
339 memcpy(a, ident, len+1);
340 a[itr] = correct_alpha[jtr];
341 array[index + ktr] = a;
348 static size_t correct_insertion(const char *ident, char **array, size_t index) {
352 const size_t len = strlen(ident);
354 for (; itr <= len; itr++) {
355 for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
356 char *a = (char*)correct_pool_alloc(len+2);
357 memcpy(a, ident, itr);
358 memcpy(a + itr + 1, ident + itr, len - itr + 1);
359 a[itr] = correct_alpha[jtr];
360 array[index + ktr] = a;
367 static GMQCC_INLINE size_t correct_size(const char *ident) {
370 * transposition = len - 1
371 * alteration = len * sizeof(correct_alpha)
372 * insertion = (len + 1) * sizeof(correct_alpha)
375 register size_t len = strlen(ident);
376 return (len) + (len - 1) + (len * (sizeof(correct_alpha)-1)) + ((len + 1) * (sizeof(correct_alpha)-1));
379 static char **correct_edit(const char *ident) {
381 char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
386 next = correct_deletion (ident, find, 0);
387 next += correct_transposition(ident, find, next);
388 next += correct_alteration (ident, find, next);
389 /*****/ correct_insertion (ident, find, next);
395 * We could use a hashtable but the space complexity isn't worth it
396 * since we're only going to determine the "did you mean?" identifier
399 static int correct_exist(char **array, size_t rows, char *ident) {
401 for (itr = 0; itr < rows; itr++)
402 if (!memcmp(array[itr], ident, strlen(ident)))
408 static GMQCC_INLINE char **correct_known_resize(char **res, size_t *allocated, size_t size) {
409 size_t oldallocated = *allocated;
411 if (size+1 < *allocated)
415 out = correct_pool_alloc(sizeof(*res) * *allocated);
416 memcpy(out, res, sizeof(*res) * oldallocated);
420 static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) {
426 char **res = correct_pool_alloc(sizeof(char *) * nxt);
429 for (; itr < rows; itr++) {
430 end = correct_edit(array[itr]);
431 row = correct_size(array[itr]);
433 for (; jtr < row; jtr++) {
434 if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
435 res = correct_known_resize(res, &nxt, len+1);
436 res[len++] = end[jtr];
445 static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
451 for (; itr < rows; itr++) {
452 if ((itm = correct_find(table, array[itr])) && (*itm > top)) {
462 * This is the exposed interface:
463 * takes a table for the dictonary a vector of sizes (used for internal
464 * probability calculation, and an identifier to "correct"
466 * the add function works the same. Except the identifier is used to
467 * add to the dictonary.
469 char *correct_str(correct_trie_t* table, const char *ident) {
472 char *e1ident = NULL;
473 char *e2ident = NULL;
479 /* needs to be allocated for free later */
480 if (correct_find(table, ident))
481 return correct_pool_claim(ident);
483 if ((e1rows = correct_size(ident))) {
484 e1 = correct_edit(ident);
486 if ((e1ident = correct_maximum(table, e1, e1rows)))
487 return correct_pool_claim(e1ident);
490 e2 = correct_known(table, e1, e1rows, &e2rows);
491 if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows))))
492 return correct_pool_claim(e2ident);
495 correct_pool_delete();
496 return util_strdup(ident);