/*
- * Copyright (C) 2012, 2013
+ * Copyright (C) 2012, 2013, 2014
* Dale Weiler
+ * Wolfgang Bumiller
*
* Permission is hereby granted, free of charge, to any person obtaining a copy of
* this software and associated documentation files (the "Software"), to deal in
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
+#include <string.h>
#include "gmqcc.h"
/*
* out of all possible corrections that maximizes the probability of C
* for the original identifer I.
*
- * Bayes' Therom suggests something of the following:
+ * Thankfully there exists some theroies for probalistic interpretations
+ * of data. Since we're operating on two distictive intepretations, the
+ * transposition from I to C. We need something that can express how much
+ * degree of I should rationally change to become C. this is called the
+ * Bayesian interpretation. You can read more about it from here:
+ * http://www.celiagreen.com/charlesmccreery/statistics/bayestutorial.pdf
+ * (which is probably the only good online documentation for bayes theroy
+ * no lie. Everything else just sucks ..)
+ *
+ * Bayes' Thereom suggests something like the following:
* AC P(I|C) P(C) / P(I)
- * Since P(I) is the same for every possibly I, we can ignore it giving
+ *
+ * However since P(I) is the same for every possibility of I, we can
+ * completley ignore it giving just:
* AC P(I|C) P(C)
*
* This greatly helps visualize how the parts of the expression are performed
* 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
* 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.
- *
+ * I had concieved of checking everything.
+ *
* A little information on additional algorithms used:
*
* Initially when I implemented this corrector, it was very slow.
* 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%
+ * down the runtime by a factor > 100%
+ *
+ * The trie initially was designed to work on all strings, but later it
+ * became aparent that not only was this not a requirement. It was also
+ * slowing down get/sets' for the trie. To fully understand, only
+ * correct_alpha needs to be understood by the trie system, knowing this
+ * We can combat the slowness using a very clever but evil optimization.
+ * By Setting a fixed sized amount of branches for the trie using a
+ * char-to-index map into the branches. We've complelty made the trie
+ * accesses entierly constant in lookup time. No really, a lookup is
+ * literally trie[str[0]] [str[1]] [2] .... .value.
+ *
*
* Future Work (If we really need it)
*
- * Currently we can only distinguishes one source of error in the
+ * Currently we can only distinguish 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.
+ * this could become a problem when lots of identifiers are involved.
*/
-#define CORRECT_POOLSIZE (128*1024*1024)
+#define CORRECT_POOL_SIZE (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);
+ correct_pool_this = (unsigned char *)mem_a(CORRECT_POOL_SIZE);
vec_push(correct_pool_data, correct_pool_this);
}
static GMQCC_INLINE void *correct_pool_alloc(size_t bytes) {
void *data;
- if (correct_pool_addr + bytes >= CORRECT_POOLSIZE)
+ if (correct_pool_addr + bytes>= CORRECT_POOL_SIZE)
correct_pool_new();
- data = correct_pool_this;
+ data = (void*)correct_pool_this;
correct_pool_this += bytes;
correct_pool_addr += bytes;
-
return data;
}
static GMQCC_INLINE char *correct_pool_claim(const char *data) {
char *claim = util_strdup(data);
- correct_pool_delete();
return claim;
}
/*
- * A fast space efficent trie for a disctonary of identifiers. This is
+ * _ 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 ... */
+
+static const size_t correct_alpha_index[0x80] = {
+ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+ 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 0, 0, 0, 0, 52,
+ 0, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40,
+ 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 0, 0, 0, 0, 0
+};
+
+/*
+ * A fast space efficent trie for a dictionary 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.
- */
+ * a hashtable becomes slow.
+ */
correct_trie_t* correct_trie_new() {
correct_trie_t *t = (correct_trie_t*)mem_a(sizeof(correct_trie_t));
t->value = NULL;
return t;
}
-void correct_trie_del_sub(correct_trie_t *t) {
+static GMQCC_INLINE void correct_trie_del_sub(correct_trie_t *t) {
size_t i;
- for (i = 0; i < vec_size(t->entries); ++i)
+ if (!t->entries)
+ return;
+ for (i = 0; i < sizeof(correct_alpha)-1; ++i) {
correct_trie_del_sub(&t->entries[i]);
- vec_free(t->entries);
+ }
+ mem_d(t->entries);
}
-void correct_trie_del(correct_trie_t *t) {
+static GMQCC_INLINE 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);
+ if (t->entries) {
+ for (i = 0; i < sizeof(correct_alpha)-1; ++i)
+ correct_trie_del_sub(&t->entries[i]);
+ mem_d(t->entries);
+ }
mem_d(t);
}
-void* correct_trie_get(const correct_trie_t *t, const char *key) {
+static GMQCC_INLINE 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)
+ if (!t->entries)
return NULL;
+ t = t->entries + correct_alpha_index[*data];
+ ++data;
}
return t->value;
}
-void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
+static GMQCC_INLINE 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);
-
- elem->ch = ch;
- elem->value = NULL;
- elem->entries = NULL;
- t = elem;
+ if (!t->entries) {
+ t->entries = (correct_trie_t*)mem_a(sizeof(correct_trie_t)*(sizeof(correct_alpha)-1));
+ memset(t->entries, 0, sizeof(correct_trie_t)*(sizeof(correct_alpha)-1));
}
+ t = t->entries + correct_alpha_index[*data];
++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) {
+ */
+static GMQCC_INLINE 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) {
+static GMQCC_INLINE bool correct_update(correct_trie_t* *table, const char *word) {
size_t *data = correct_find(*table, word);
if (!data)
- return 0;
+ return false;
(*data)++;
- return 1;
+ return true;
}
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;
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
* 2) transposition
* 3) alteration
* 4) insertion
+ *
+ * These functions could take an additional size_t **size paramater
+ * and store back the results of their new length in an array that
+ * is the same as **array for the memcmp in correct_exists. I'm just
+ * not able to figure out how to do that just yet. As my brain is
+ * not in the mood to figure out that logic. This is a reminder to
+ * do it, or for someone else to :-) correct_edit however would also
+ * need to take a size_t ** to carry it along (would all the argument
+ * overhead be worth it?)
*/
-static size_t correct_deletion(const char *ident, char **array, size_t index) {
+static GMQCC_INLINE size_t correct_deletion(const char *ident, char **array) {
size_t itr = 0;
const size_t len = strlen(ident);
char *a = (char*)correct_pool_alloc(len+1);
memcpy(a, ident, itr);
memcpy(a + itr, ident + itr + 1, len - itr);
- array[index + itr] = a;
+ array[itr] = a;
}
return itr;
}
-static size_t correct_transposition(const char *ident, char **array, size_t index) {
+static GMQCC_INLINE size_t correct_transposition(const char *ident, char **array) {
size_t itr = 0;
const size_t len = strlen(ident);
tmp = a[itr];
a[itr ] = a[itr+1];
a[itr+1] = tmp;
- array[index + itr] = a;
+ array[itr] = a;
}
return itr;
}
-static size_t correct_alteration(const char *ident, char **array, size_t index) {
+static GMQCC_INLINE size_t correct_alteration(const char *ident, char **array) {
size_t itr = 0;
size_t jtr = 0;
size_t ktr = 0;
char *a = (char*)correct_pool_alloc(len+1);
memcpy(a, ident, len+1);
a[itr] = correct_alpha[jtr];
- array[index + ktr] = a;
+ array[ktr] = a;
}
}
return ktr;
}
-static size_t correct_insertion(const char *ident, char **array, size_t index) {
+static GMQCC_INLINE size_t correct_insertion(const char *ident, char **array) {
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++) {
+ for (jtr = 0; jtr < sizeof(correct_alpha)-1; jtr++) {
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;
+ array[itr * (sizeof(correct_alpha)-1) + jtr] = a;
}
}
- return ktr;
+ return (len+1)*(sizeof(correct_alpha)-1);
}
static GMQCC_INLINE size_t correct_size(const char *ident) {
* transposition = len - 1
* alteration = len * sizeof(correct_alpha)
* insertion = (len + 1) * sizeof(correct_alpha)
- */
+ */
register size_t len = strlen(ident);
return (len) + (len - 1) + (len * (sizeof(correct_alpha)-1)) + ((len + 1) * (sizeof(correct_alpha)-1));
}
-static char **correct_edit(const char *ident) {
+static GMQCC_INLINE char **correct_edit(const char *ident, size_t **lens) {
size_t next;
- char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
+ size_t size = correct_size(ident);
+ char **find = (char**)correct_pool_alloc(size * sizeof(char*));
- if (!find)
+ if (!find || !(*lens = (size_t*)correct_pool_alloc(size * sizeof(size_t))))
return NULL;
- next = correct_deletion (ident, find, 0);
- next += correct_transposition(ident, find, next);
- next += correct_alteration (ident, find, next);
- /*****/ correct_insertion (ident, find, next);
+ next = correct_deletion (ident, find);
+ next += correct_transposition(ident, find+next);
+ next += correct_alteration (ident, find+next);
+ /*****/ correct_insertion (ident, find+next);
+
+ /* precompute lengths */
+ for (next = 0; next < size; next++)
+ (*lens)[next] = strlen(find[next]);
return find;
}
-/*
- * We could use a hashtable but the space complexity isn't worth it
- * since we're only going to determine the "did you mean?" identifier
- * on error.
- */
-static int correct_exist(char **array, size_t rows, char *ident) {
+static GMQCC_INLINE int correct_exist(char **array, register size_t *sizes, size_t rows, char *ident, register size_t len) {
size_t itr;
- for (itr = 0; itr < rows; itr++)
- if (!strcmp(array[itr], ident))
+ for (itr = 0; itr < rows; itr++) {
+ /*
+ * We can save tons of calls to memcmp if we simply ignore comparisions
+ * that we know cannot contain the same length.
+ */
+ if (sizes[itr] == len && !memcmp(array[itr], ident, len))
return 1;
+ }
return 0;
}
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)
+ if (size < oldallocated)
return res;
- *allocated += 32;
- out = correct_pool_alloc(sizeof(*res) * *allocated);
+ out = (char**)correct_pool_alloc(sizeof(*res) * oldallocated + 32);
memcpy(out, res, sizeof(*res) * oldallocated);
+
+ *allocated += 32;
return out;
}
-static char **correct_known(correct_trie_t* table, char **array, size_t rows, size_t *next) {
+static char **correct_known(correction_t *corr, 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;
+ char **res = (char**)correct_pool_alloc(sizeof(char *) * nxt);
+ char **end = NULL;
+ size_t *bit = NULL;
for (; itr < rows; itr++) {
- end = correct_edit(array[itr]);
+ if (!array[itr][0])
+ continue;
+ if (vec_size(corr->edits) > itr+1) {
+ end = corr->edits[itr+1];
+ bit = corr->lens [itr+1];
+ } else {
+ end = correct_edit(array[itr], &bit);
+ vec_push(corr->edits, end);
+ vec_push(corr->lens, bit);
+ }
row = correct_size(array[itr]);
- for (; jtr < row; jtr++) {
- if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
+ for (jtr = 0; jtr < row; jtr++) {
+ if (correct_find(table, end[jtr]) && !correct_exist(res, bit, len, end[jtr], bit[jtr])) {
res = correct_known_resize(res, &nxt, len+1);
res[len++] = end[jtr];
}
return res;
}
-static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
+static GMQCC_INLINE char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
char *str = NULL;
size_t *itm = NULL;
size_t itr = 0;
/*
* This is the exposed interface:
* takes a table for the dictonary a vector of sizes (used for internal
- * probability calculation, and an identifier to "correct"
- *
- * the add function works the same. Except the identifier is used to
- * add to the dictonary.
+ * probability calculation), and an identifier to "correct".
*/
-char *correct_str(correct_trie_t* table, const char *ident) {
+void correct_init(correction_t *c)
+{
+ correct_pool_new();
+ c->edits = NULL;
+ c->lens = NULL;
+}
+
+void correct_free(correction_t *c)
+{
+ vec_free(c->edits);
+ vec_free(c->lens);
+ correct_pool_delete();
+}
+
+char *correct_str(correction_t *corr, correct_trie_t* table, const char *ident) {
char **e1 = NULL;
char **e2 = NULL;
char *e1ident = NULL;
char *e2ident = NULL;
size_t e1rows = 0;
size_t e2rows = 0;
-
- correct_pool_new();
+ size_t *bits = NULL;
/* needs to be allocated for free later */
if (correct_find(table, ident))
return correct_pool_claim(ident);
if ((e1rows = correct_size(ident))) {
- e1 = correct_edit(ident);
+ if (vec_size(corr->edits) > 0)
+ e1 = corr->edits[0];
+ else {
+ e1 = correct_edit(ident, &bits);
+ vec_push(corr->edits, e1);
+ vec_push(corr->lens, bits);
+ }
if ((e1ident = correct_maximum(table, e1, e1rows)))
return correct_pool_claim(e1ident);
}
- e2 = correct_known(table, e1, e1rows, &e2rows);
+ e2 = correct_known(corr, table, e1, e1rows, &e2rows);
if (e2rows && ((e2ident = correct_maximum(table, e2, e2rows))))
return correct_pool_claim(e2ident);
- correct_pool_delete();
return util_strdup(ident);
}