2 * Copyright (C) 2012, 2013
6 * Permission is hereby granted, free of charge, to any person obtaining a copy of
7 * this software and associated documentation files (the "Software"), to deal in
8 * the Software without restriction, including without limitation the rights to
9 * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
10 * of the Software, and to permit persons to whom the Software is furnished to do
11 * so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
27 * This is a very clever method for correcting mistakes in QuakeC code
28 * most notably when invalid identifiers are used or inproper assignments;
29 * we can proprly lookup in multiple dictonaries (depening on the rules
30 * of what the task is trying to acomplish) to find the best possible
34 * A little about how it works, and probability theory:
36 * When given an identifier (which we will denote I), we're essentially
37 * just trying to choose the most likely correction for that identifier.
38 * (the actual "correction" can very well be the identifier itself).
39 * There is actually no way to know for sure that certian identifers
40 * such as "lates", need to be corrected to "late" or "latest" or any
41 * other permutations that look lexically the same. This is why we
42 * must advocate the usage of probabilities. This means that instead of
43 * just guessing, instead we're trying to find the correction for C,
44 * out of all possible corrections that maximizes the probability of C
45 * for the original identifer I.
47 * Bayes' Therom suggests something of the following:
48 * AC P(I|C) P(C) / P(I)
49 * Since P(I) is the same for every possibly I, we can ignore it giving
52 * This greatly helps visualize how the parts of the expression are performed
53 * there is essentially three, from right to left we perform the following:
55 * 1: P(C), the probability that a proposed correction C will stand on its
56 * own. This is called the language model.
58 * 2: P(I|C), the probability that I would be used, when the programmer
59 * really meant C. This is the error model.
61 * 3: AC, the control mechanisim, an enumerator if you will, one that
62 * enumerates all feasible values of C, to determine the one that
63 * gives the greatest probability score.
65 * In reality the requirement for a more complex expression involving
66 * two seperate models is considerably a waste. But one must recognize
67 * that P(C|I) is already conflating two factors. It's just much simpler
68 * to seperate the two models and deal with them explicitaly. To properly
69 * estimate P(C|I) you have to consider both the probability of C and
70 * probability of the transposition from C to I. It's simply much more
71 * cleaner, and direct to seperate the two factors.
73 * Research tells us that 80% to 95% of all spelling errors have an edit
74 * distance no greater than one. Knowing this we can optimize for most
75 * cases of mistakes without taking a performance hit. Which is what we
76 * base longer edit distances off of. Opposed to the original method of
77 * I had concieved of checking everything.
79 * A little information on additional algorithms used:
81 * Initially when I implemented this corrector, it was very slow.
82 * Need I remind you this is essentially a brute force attack on strings,
83 * and since every transformation requires dynamic memory allocations,
84 * you can easily imagine where most of the runtime conflated. Yes
85 * It went right to malloc. More than THREE MILLION malloc calls are
86 * performed for an identifier about 16 bytes long. This was such a
87 * shock to me. A forward allocator (or as some call it a bump-point
88 * allocator, or just a memory pool) was implemented. To combat this.
90 * But of course even other factors were making it slow. Initially
91 * this used a hashtable. And hashtables have a good constant lookup
92 * time complexity. But the problem wasn't in the hashtable, it was
93 * in the hashing (despite having one of the fastest hash functions
94 * known). Remember those 3 million mallocs? Well for every malloc
95 * there is also a hash. After 3 million hashes .. you start to get
96 * very slow. To combat this I had suggested burst tries to Blub.
97 * The next day he had implemented them. Sure enough this brought
98 * down the runtime by a factory > 100%
100 * Future Work (If we really need it)
102 * Currently we can only distinguishes one source of error in the
103 * language model we use. This could become an issue for identifiers
104 * that have close colliding rates, e.g colate->coat yields collate.
106 * Currently the error model has been fairly trivial, the smaller the
107 * edit distance the smaller the error. This usually causes some un-
108 * expected problems. e.g reciet->recite yields recipt. For QuakeC
109 * this could become a problem when lots of identifiers are involved.
111 * Our control mechanisim could use a limit, i.e limit the number of
112 * sets of edits for distance X. This would also increase execution
113 * speed considerably.
117 #define CORRECT_POOLSIZE (128*1024*1024)
119 * A forward allcator for the corrector. This corrector requires a lot
120 * of allocations. This forward allocator combats all those allocations
121 * and speeds us up a little. It also saves us space in a way since each
122 * allocation isn't wasting a little header space for when NOTRACK isn't
125 static unsigned char **correct_pool_data = NULL;
126 static unsigned char *correct_pool_this = NULL;
127 static size_t correct_pool_addr = 0;
129 static GMQCC_INLINE void correct_pool_new(void) {
130 correct_pool_addr = 0;
131 correct_pool_this = (unsigned char *)mem_a(CORRECT_POOLSIZE);
133 vec_push(correct_pool_data, correct_pool_this);
136 static GMQCC_INLINE void *correct_pool_alloc(size_t bytes) {
138 if (correct_pool_addr + bytes >= CORRECT_POOLSIZE)
141 data = correct_pool_this;
142 correct_pool_this += bytes;
143 correct_pool_addr += bytes;
148 static GMQCC_INLINE void correct_pool_delete(void) {
150 for (i = 0; i < vec_size(correct_pool_data); ++i)
151 mem_d(correct_pool_data[i]);
153 correct_pool_data = NULL;
154 correct_pool_this = NULL;
155 correct_pool_addr = 0;
159 static GMQCC_INLINE char *correct_pool_claim(const char *data) {
160 char *claim = util_strdup(data);
161 correct_pool_delete();
166 * A fast space efficent trie for a disctonary of identifiers. This is
167 * faster than a hashtable for one reason. A hashtable itself may have
168 * fast constant lookup time, but the hash itself must be very fast. We
169 * have one of the fastest hash functions for strings, but if you do a
170 * lost of hashing (which we do, almost 3 million hashes per identifier)
171 * a hashtable becomes slow. Very Very Slow.
173 correct_trie_t* correct_trie_new() {
174 correct_trie_t *t = (correct_trie_t*)mem_a(sizeof(correct_trie_t));
180 void correct_trie_del_sub(correct_trie_t *t) {
182 for (i = 0; i < vec_size(t->entries); ++i)
183 correct_trie_del_sub(&t->entries[i]);
184 vec_free(t->entries);
187 void correct_trie_del(correct_trie_t *t) {
189 for (i = 0; i < vec_size(t->entries); ++i)
190 correct_trie_del_sub(&t->entries[i]);
191 vec_free(t->entries);
195 void* correct_trie_get(const correct_trie_t *t, const char *key) {
196 const unsigned char *data = (const unsigned char*)key;
198 unsigned char ch = *data;
199 const size_t vs = vec_size(t->entries);
201 const correct_trie_t *entries = t->entries;
202 for (i = 0; i < vs; ++i) {
203 if (entries[i].ch == ch) {
215 void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
216 const unsigned char *data = (const unsigned char*)key;
218 const size_t vs = vec_size(t->entries);
219 unsigned char ch = *data;
220 correct_trie_t *entries = t->entries;
223 for (i = 0; i < vs; ++i) {
224 if (entries[i].ch == ch) {
230 correct_trie_t *elem = (correct_trie_t*)vec_add(t->entries, 1);
234 elem->entries = NULL;
244 * Implementation of the corrector algorithm commences. A very efficent
245 * brute-force attack (thanks to tries and mempool :-)).
247 static size_t *correct_find(correct_trie_t *table, const char *word) {
248 return (size_t*)correct_trie_get(table, word);
251 static int correct_update(correct_trie_t* *table, const char *word) {
252 size_t *data = correct_find(*table, word);
260 void correct_add(correct_trie_t* table, size_t ***size, const char *ident) {
262 const char *add = ident;
264 if (!correct_update(&table, add)) {
265 data = (size_t*)mem_a(sizeof(size_t));
268 vec_push((*size), data);
269 correct_trie_set(table, add, data);
273 void correct_del(correct_trie_t* dictonary, size_t **data) {
275 const size_t vs = vec_size(data);
277 for (i = 0; i < vs; i++)
281 correct_trie_del(dictonary);
285 * _ is valid in identifiers. I've yet to implement numerics however
286 * because they're only valid after the first character is of a _, or
289 static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyz"
290 "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
291 "_"; /* TODO: Numbers ... */
294 * correcting logic for the following forms of transformations:
300 * These functions could take an additional size_t **size paramater
301 * and store back the results of their new length in an array that
302 * is the same as **array for the memcmp in correct_exists. I'm just
303 * not able to figure out how to do that just yet. As my brain is
304 * not in the mood to figure out that logic. This is a reminder to
305 * do it, or for someone else to :-) correct_edit however would also
306 * need to take a size_t ** to carry it along (would all the argument
307 * overhead be worth it?)
309 static size_t correct_deletion(const char *ident, char **array, size_t index) {
311 const size_t len = strlen(ident);
313 for (; itr < len; itr++) {
314 char *a = (char*)correct_pool_alloc(len+1);
315 memcpy(a, ident, itr);
316 memcpy(a + itr, ident + itr + 1, len - itr);
317 array[index + itr] = a;
323 static size_t correct_transposition(const char *ident, char **array, size_t index) {
325 const size_t len = strlen(ident);
327 for (; itr < len - 1; itr++) {
329 char *a = (char*)correct_pool_alloc(len+1);
330 memcpy(a, ident, len+1);
334 array[index + itr] = a;
340 static size_t correct_alteration(const char *ident, char **array, size_t index) {
344 const size_t len = strlen(ident);
346 for (; itr < len; itr++) {
347 for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
348 char *a = (char*)correct_pool_alloc(len+1);
349 memcpy(a, ident, len+1);
350 a[itr] = correct_alpha[jtr];
351 array[index + ktr] = a;
358 static size_t correct_insertion(const char *ident, char **array, size_t index) {
362 const size_t len = strlen(ident);
364 for (; itr <= len; itr++) {
365 for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
366 char *a = (char*)correct_pool_alloc(len+2);
367 memcpy(a, ident, itr);
368 memcpy(a + itr + 1, ident + itr, len - itr + 1);
369 a[itr] = correct_alpha[jtr];
370 array[index + ktr] = a;
377 static GMQCC_INLINE size_t correct_size(const char *ident) {
380 * transposition = len - 1
381 * alteration = len * sizeof(correct_alpha)
382 * insertion = (len + 1) * sizeof(correct_alpha)
385 register size_t len = strlen(ident);
386 return (len) + (len - 1) + (len * (sizeof(correct_alpha)-1)) + ((len + 1) * (sizeof(correct_alpha)-1));
389 static char **correct_edit(const char *ident) {
391 char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
396 next = correct_deletion (ident, find, 0);
397 next += correct_transposition(ident, find, next);
398 next += correct_alteration (ident, find, next);
399 /*****/ correct_insertion (ident, find, next);
405 * We could use a hashtable but the space complexity isn't worth it
406 * since we're only going to determine the "did you mean?" identifier
409 static int correct_exist(char **array, size_t rows, char *ident) {
411 for (itr = 0; itr < rows; itr++)
412 if (!memcmp(array[itr], ident, strlen(ident)))
418 static GMQCC_INLINE char **correct_known_resize(char **res, size_t *allocated, size_t size) {
419 size_t oldallocated = *allocated;
421 if (size+1 < *allocated)
425 out = correct_pool_alloc(sizeof(*res) * *allocated);
426 memcpy(out, res, sizeof(*res) * oldallocated);
430 static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) {
436 char **res = correct_pool_alloc(sizeof(char *) * nxt);
439 for (; itr < rows; itr++) {
440 end = correct_edit(array[itr]);
441 row = correct_size(array[itr]);
443 for (; jtr < row; jtr++) {
444 if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
445 res = correct_known_resize(res, &nxt, len+1);
446 res[len++] = end[jtr];
455 static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
461 for (; itr < rows; itr++) {
462 if ((itm = correct_find(table, array[itr])) && (*itm > top)) {
472 * This is the exposed interface:
473 * takes a table for the dictonary a vector of sizes (used for internal
474 * probability calculation, and an identifier to "correct"
476 * the add function works the same. Except the identifier is used to
477 * add to the dictonary.
479 char *correct_str(correct_trie_t* table, const char *ident) {
482 char *e1ident = NULL;
483 char *e2ident = NULL;
489 /* needs to be allocated for free later */
490 if (correct_find(table, ident))
491 return correct_pool_claim(ident);
493 if ((e1rows = correct_size(ident))) {
494 e1 = correct_edit(ident);
496 if ((e1ident = correct_maximum(table, e1, e1rows)))
497 return correct_pool_claim(e1ident);
500 e2 = correct_known(table, e1, e1rows, &e2rows);
501 if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows))))
502 return correct_pool_claim(e2ident);
505 correct_pool_delete();
506 return util_strdup(ident);