* There is actually no way to know for sure that certian identifers
* such as "lates", need to be corrected to "late" or "latest" or any
* other permutations that look lexically the same. This is why we
- * must advocate the usage of probabilities. This implies that we're
- * trying to find the correction for C, out of all possible corrections
- * that maximizes the probability of C for the original identifer I.
+ * must advocate the usage of probabilities. This means that instead of
+ * just guessing, instead we're trying to find the correction for C,
+ * out of all possible corrections that maximizes the probability of C
+ * for the original identifer I.
*
* Bayes' Therom suggests something of the following:
* AC P(I|C) P(C) / P(I)
* 2: P(I|C), the probability that I would be used, when the programmer
* really meant C. This is the error model.
*
- * 3: AC, the control mechanisim, which implies the enumeration of all
- * feasible values of C, and then determine the one that gives the
- * greatest probability score. Selecting it as the "correction"
- *
- *
- * The requirement for complex expression involving two models:
+ * 3: AC, the control mechanisim, an enumerator if you will, one that
+ * enumerates all feasible values of C, to determine the one that
+ * gives the greatest probability score.
*
- * In reality the requirement for a more complex expression involving
+ * In reality the requirement for a more complex expression involving
* two seperate models is considerably a waste. But one must recognize
* that P(C|I) is already conflating two factors. It's just much simpler
* to seperate the two models and deal with them explicitaly. To properly
* estimate P(C|I) you have to consider both the probability of C and
* probability of the transposition from C to I. It's simply much more
* cleaner, and direct to seperate the two factors.
+ *
+ * Research tells us that 80% to 95% of all spelling errors have an edit
+ * distance no greater than one. Knowing this we can optimize for most
+ * cases of mistakes without taking a performance hit. Which is what we
+ * base longer edit distances off of. Opposed to the original method of
+ * I had concieved of checking everything.
+ *
+ * A little information on additional algorithms used:
+ *
+ * Initially when I implemented this corrector, it was very slow.
+ * Need I remind you this is essentially a brute force attack on strings,
+ * and since every transformation requires dynamic memory allocations,
+ * you can easily imagine where most of the runtime conflated. Yes
+ * It went right to malloc. More than THREE MILLION malloc calls are
+ * performed for an identifier about 16 bytes long. This was such a
+ * shock to me. A forward allocator (or as some call it a bump-point
+ * allocator, or just a memory pool) was implemented. To combat this.
+ *
+ * But of course even other factors were making it slow. Initially
+ * this used a hashtable. And hashtables have a good constant lookup
+ * time complexity. But the problem wasn't in the hashtable, it was
+ * in the hashing (despite having one of the fastest hash functions
+ * known). Remember those 3 million mallocs? Well for every malloc
+ * there is also a hash. After 3 million hashes .. you start to get
+ * very slow. To combat this I had suggested burst tries to Blub.
+ * The next day he had implemented them. Sure enough this brought
+ * down the runtime by a factory > 100%
+ *
+ * Future Work (If we really need it)
+ *
+ * Currently we can only distinguishes one source of error in the
+ * language model we use. This could become an issue for identifiers
+ * that have close colliding rates, e.g colate->coat yields collate.
+ *
+ * Currently the error model has been fairly trivial, the smaller the
+ * edit distance the smaller the error. This usually causes some un-
+ * expected problems. e.g reciet->recite yields recipt. For QuakeC
+ * this could become a problem when lots of identifiers are involved.
+ *
+ * Our control mechanisim could use a limit, i.e limit the number of
+ * sets of edits for distance X. This would also increase execution
+ * speed considerably.
*/
-/* some hashtable management for dictonaries */
-static size_t *correct_find(ht table, const char *word) {
- return (size_t*)util_htget(table, word);
+
+#define CORRECT_POOLSIZE (128*1024*1024)
+/*
+ * A forward allcator for the corrector. This corrector requires a lot
+ * of allocations. This forward allocator combats all those allocations
+ * and speeds us up a little. It also saves us space in a way since each
+ * allocation isn't wasting a little header space for when NOTRACK isn't
+ * defined.
+ */
+static unsigned char **correct_pool_data = NULL;
+static unsigned char *correct_pool_this = NULL;
+static size_t correct_pool_addr = 0;
+
+static GMQCC_INLINE void correct_pool_new(void) {
+ correct_pool_addr = 0;
+ correct_pool_this = (unsigned char *)mem_a(CORRECT_POOLSIZE);
+
+ vec_push(correct_pool_data, correct_pool_this);
}
-static int correct_update(ht *table, const char *word) {
- size_t *data = correct_find(*table, word);
- if (!data)
- return 0;
+static GMQCC_INLINE void *correct_pool_alloc(size_t bytes) {
+ void *data;
+ if (correct_pool_addr + bytes >= CORRECT_POOLSIZE)
+ correct_pool_new();
- (*data)++;
- return 1;
+ data = correct_pool_this;
+ correct_pool_this += bytes;
+ correct_pool_addr += bytes;
+
+ return data;
}
+static GMQCC_INLINE void correct_pool_delete(void) {
+ size_t i;
+ for (i = 0; i < vec_size(correct_pool_data); ++i)
+ mem_d(correct_pool_data[i]);
+
+ correct_pool_data = NULL;
+ correct_pool_this = NULL;
+ correct_pool_addr = 0;
+}
+
+
+static GMQCC_INLINE char *correct_pool_claim(const char *data) {
+ char *claim = util_strdup(data);
+ correct_pool_delete();
+ return claim;
+}
/*
- * _ is valid in identifiers. I've yet to implement numerics however
- * because they're only valid after the first character is of a _, or
- * alpha character.
- */
-static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_";
+ * A fast space efficent trie for a disctonary of identifiers. This is
+ * faster than a hashtable for one reason. A hashtable itself may have
+ * fast constant lookup time, but the hash itself must be very fast. We
+ * have one of the fastest hash functions for strings, but if you do a
+ * lost of hashing (which we do, almost 3 million hashes per identifier)
+ * a hashtable becomes slow. Very Very Slow.
+ */
+correct_trie_t* correct_trie_new() {
+ correct_trie_t *t = (correct_trie_t*)mem_a(sizeof(correct_trie_t));
+ t->value = NULL;
+ t->entries = NULL;
+ return t;
+}
-static char *correct_strndup(const char *src, size_t n) {
- char *ret;
- size_t len = strlen(src);
+void correct_trie_del_sub(correct_trie_t *t) {
+ size_t i;
+ for (i = 0; i < vec_size(t->entries); ++i)
+ correct_trie_del_sub(&t->entries[i]);
+ vec_free(t->entries);
+}
- if (n < len)
- len = n;
+void correct_trie_del(correct_trie_t *t) {
+ size_t i;
+ for (i = 0; i < vec_size(t->entries); ++i)
+ correct_trie_del_sub(&t->entries[i]);
+ vec_free(t->entries);
+ mem_d(t);
+}
- if (!(ret = (char*)mem_a(len + 1)))
- return NULL;
+void* correct_trie_get(const correct_trie_t *t, const char *key) {
+ const unsigned char *data = (const unsigned char*)key;
+ while (*data) {
+ unsigned char ch = *data;
+ const size_t vs = vec_size(t->entries);
+ size_t i;
+ const correct_trie_t *entries = t->entries;
+ for (i = 0; i < vs; ++i) {
+ if (entries[i].ch == ch) {
+ t = &entries[i];
+ ++data;
+ break;
+ }
+ }
+ if (i == vs)
+ return NULL;
+ }
+ return t->value;
+}
+
+void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
+ const unsigned char *data = (const unsigned char*)key;
+ while (*data) {
+ const size_t vs = vec_size(t->entries);
+ unsigned char ch = *data;
+ correct_trie_t *entries = t->entries;
+ size_t i;
+
+ for (i = 0; i < vs; ++i) {
+ if (entries[i].ch == ch) {
+ t = &entries[i];
+ break;
+ }
+ }
+ if (i == vs) {
+ correct_trie_t *elem = (correct_trie_t*)vec_add(t->entries, 1);
- ret[len] = '\0';
- return (char*)memcpy(ret, src, len);
+ elem->ch = ch;
+ elem->value = NULL;
+ elem->entries = NULL;
+ t = elem;
+ }
+ ++data;
+ }
+ t->value = value;
+}
+
+
+/*
+ * Implementation of the corrector algorithm commences. A very efficent
+ * brute-force attack (thanks to tries and mempool :-)).
+ */
+static size_t *correct_find(correct_trie_t *table, const char *word) {
+ return (size_t*)correct_trie_get(table, word);
+}
+
+static int correct_update(correct_trie_t* *table, const char *word) {
+ size_t *data = correct_find(*table, word);
+ if (!data)
+ return 0;
+
+ (*data)++;
+ return 1;
}
-static char *correct_concat(char *str1, char *str2, bool next) {
- char *ret = NULL;
+void correct_add(correct_trie_t* table, size_t ***size, const char *ident) {
+ size_t *data = NULL;
+ const char *add = ident;
+
+ if (!correct_update(&table, add)) {
+ data = (size_t*)mem_a(sizeof(size_t));
+ *data = 1;
-#if 0
- if (!str1) {
- str1 = mem_a(1);
- *str1 = '\0';
+ vec_push((*size), data);
+ correct_trie_set(table, add, data);
}
-#endif
+}
- str1 = mem_r (str1, strlen(str1) + strlen(str2) + 1);
- ret = strcat(str1, str2);
+void correct_del(correct_trie_t* dictonary, size_t **data) {
+ size_t i;
+ const size_t vs = vec_size(data);
- if (str2 && next)
- mem_d(str2);
+ for (i = 0; i < vs; i++)
+ mem_d(data[i]);
- return ret;
+ vec_free(data);
+ correct_trie_del(dictonary);
}
+/*
+ * _ is valid in identifiers. I've yet to implement numerics however
+ * because they're only valid after the first character is of a _, or
+ * alpha character.
+ */
+static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyz"
+ "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+ "_"; /* TODO: Numbers ... */
+
/*
* correcting logic for the following forms of transformations:
* 1) deletion
* 4) insertion
*/
static size_t correct_deletion(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t len = strlen(ident);
-
- for (itr = 0; itr < len; itr++) {
- array[index + itr] = correct_concat (
- correct_strndup (ident, itr),
- correct_strndup (ident+itr+1, len-(itr+1)),
- true
- );
+ size_t itr = 0;
+ const size_t len = strlen(ident);
+
+ for (; itr < len; itr++) {
+ char *a = (char*)correct_pool_alloc(len+1);
+ memcpy(a, ident, itr);
+ memcpy(a + itr, ident + itr + 1, len - itr);
+ array[index + itr] = a;
}
return itr;
}
static size_t correct_transposition(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t len = strlen(ident);
-
- for (itr = 0; itr < len - 1; itr++) {
- array[index + itr] = correct_concat (
- correct_concat (
- correct_strndup(ident, itr),
- correct_strndup(ident+itr+1, 1),
- true
- ),
- correct_concat (
- correct_strndup(ident+itr, 1),
- correct_strndup(ident+itr+2, len-(itr+2)),
- true
- ),
- true
- );
+ size_t itr = 0;
+ const size_t len = strlen(ident);
+
+ for (; itr < len - 1; itr++) {
+ char tmp;
+ char *a = (char*)correct_pool_alloc(len+1);
+ memcpy(a, ident, len+1);
+ tmp = a[itr];
+ a[itr ] = a[itr+1];
+ a[itr+1] = tmp;
+ array[index + itr] = a;
}
return itr;
}
static size_t correct_alteration(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t jtr;
- size_t ktr;
- size_t len = strlen(ident);
- char cct[2] = { 0, 0 }; /* char code table, for concatenation */
-
- for (itr = 0, ktr = 0; itr < len; itr++) {
- for (jtr = 0; jtr < sizeof(correct_alpha); jtr++, ktr++) {
- *cct = correct_alpha[jtr];
- array[index + ktr] = correct_concat (
- correct_concat (
- correct_strndup(ident, itr),
- (char *) &cct,
- false
- ),
- correct_strndup (
- ident + (itr+1),
- len - (itr+1)
- ),
- true
- );
+ size_t itr = 0;
+ size_t jtr = 0;
+ size_t ktr = 0;
+ const size_t len = strlen(ident);
+
+ for (; itr < len; itr++) {
+ for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
+ char *a = (char*)correct_pool_alloc(len+1);
+ memcpy(a, ident, len+1);
+ a[itr] = correct_alpha[jtr];
+ array[index + ktr] = a;
}
}
}
static size_t correct_insertion(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t jtr;
- size_t ktr;
- size_t len = strlen(ident);
- char cct[2] = { 0, 0 }; /* char code table, for concatenation */
-
- for (itr = 0, ktr = 0; itr <= len; itr++) {
- for (jtr = 0; jtr < sizeof(correct_alpha); jtr++, ktr++) {
- *cct = correct_alpha[jtr];
- array[index + ktr] = correct_concat (
- correct_concat (
- correct_strndup (ident, itr),
- (char *) &cct,
- false
- ),
- correct_strndup (
- ident+itr,
- len - itr
- ),
- true
- );
+ size_t itr = 0;
+ size_t jtr = 0;
+ size_t ktr = 0;
+ const size_t len = strlen(ident);
+
+ for (; itr <= len; itr++) {
+ for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++, ktr++) {
+ char *a = (char*)correct_pool_alloc(len+2);
+ memcpy(a, ident, itr);
+ memcpy(a + itr + 1, ident + itr, len - itr + 1);
+ a[itr] = correct_alpha[jtr];
+ array[index + ktr] = a;
}
}
*/
register size_t len = strlen(ident);
- return (len) + (len - 1) + (len * sizeof(correct_alpha)) + ((len + 1) * sizeof(correct_alpha));
+ return (len) + (len - 1) + (len * (sizeof(correct_alpha)-1)) + ((len + 1) * (sizeof(correct_alpha)-1));
}
static char **correct_edit(const char *ident) {
size_t next;
- char **find = (char**)mem_a(correct_size(ident) * sizeof(char*));
+ char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
if (!find)
return NULL;
return 0;
}
-static char **correct_known(ht table, char **array, size_t rows, size_t *next) {
- size_t itr;
- size_t jtr;
- size_t len;
- size_t row;
- char **res = NULL;
- char **end;
+static GMQCC_INLINE char **correct_known_resize(char **res, size_t *allocated, size_t size) {
+ size_t oldallocated = *allocated;
+ char **out;
+ if (size+1 < *allocated)
+ return res;
+
+ *allocated += 32;
+ out = correct_pool_alloc(sizeof(*res) * *allocated);
+ memcpy(out, res, sizeof(*res) * oldallocated);
+ return out;
+}
- for (itr = 0, len = 0; itr < rows; itr++) {
+static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) {
+ size_t itr = 0;
+ size_t jtr = 0;
+ size_t len = 0;
+ size_t row = 0;
+ size_t nxt = 8;
+ char **res = correct_pool_alloc(sizeof(char *) * nxt);
+ char **end = NULL;
+
+ for (; itr < rows; itr++) {
end = correct_edit(array[itr]);
row = correct_size(array[itr]);
- for (jtr = 0; jtr < row; jtr++) {
+ for (; jtr < row; jtr++) {
if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
- res = mem_r(res, sizeof(char*) * (len + 1));
+ res = correct_known_resize(res, &nxt, len+1);
res[len++] = end[jtr];
- } else {
- mem_d(end[jtr]);
}
}
-
- mem_d(end);
}
*next = len;
return res;
}
-static char *correct_maximum(ht table, char **array, size_t rows) {
- char *str = NULL;
- size_t *itm = NULL;
- size_t itr;
- size_t top;
+static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
+ char *str = NULL;
+ size_t *itm = NULL;
+ size_t itr = 0;
+ size_t top = 0;
- for (itr = 0, top = 0; itr < rows; itr++) {
+ for (; itr < rows; itr++) {
if ((itm = correct_find(table, array[itr])) && (*itm > top)) {
top = *itm;
str = array[itr];
return str;
}
-static void correct_cleanup(char **array, size_t rows) {
- size_t itr;
- for (itr = 0; itr < rows; itr++)
- mem_d(array[itr]);
-
- mem_d(array);
-}
-
/*
* This is the exposed interface:
* takes a table for the dictonary a vector of sizes (used for internal
*
* the add function works the same. Except the identifier is used to
* add to the dictonary.
- */
-void correct_add(ht table, size_t ***size, const char *ident) {
- size_t *data = NULL;
- const char *add = ident;
-
- if (!correct_update(&table, add)) {
- data = (size_t*)mem_a(sizeof(size_t));
- *data = 1;
-
- vec_push((*size), data);
- util_htset(table, add, data);
- }
-}
-
-char *correct_str(ht table, const char *ident) {
+ */
+char *correct_str(correct_trie_t* table, const char *ident) {
char **e1;
char **e2;
char *e1ident;
char *e2ident;
- char *found = util_strdup(ident);
size_t e1rows = 0;
size_t e2rows = 0;
+ correct_pool_new();
+
/* needs to be allocated for free later */
if (correct_find(table, ident))
- return found;
+ return correct_pool_claim(ident);
if ((e1rows = correct_size(ident))) {
e1 = correct_edit(ident);
- if ((e1ident = correct_maximum(table, e1, e1rows))) {
- mem_d(found);
- found = util_strdup(e1ident);
- correct_cleanup(e1, e1rows);
- return found;
- }
+ if ((e1ident = correct_maximum(table, e1, e1rows)))
+ return correct_pool_claim(e1ident);
}
e2 = correct_known(table, e1, e1rows, &e2rows);
- if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows)))) {
- mem_d(found);
- found = util_strdup(e2ident);
- }
-
- correct_cleanup(e1, e1rows);
- correct_cleanup(e2, e2rows);
-
- return found;
-}
+ if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows))))
+ return correct_pool_claim(e2ident);
-void correct_del(ht dictonary, size_t **data) {
- size_t i;
- for (i = 0; i < vec_size(data); i++)
- mem_d(data[i]);
- vec_free(data);
- util_htdel(dictonary);
+ correct_pool_delete();
+ return util_strdup(ident);
}
projects / xonotic / gmqcc.git / blobdiff