/*
* Copyright (C) 2012, 2013
* 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
* the Software without restriction, including without limitation the rights to
* 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
* 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
* 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,
* 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.
*/
-#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
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;
}
}
/*
- * A fast space efficent trie for a disctonary of identifiers. This is
+ * 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));
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;
+ unsigned char ch = *data;
+ const size_t vs = vec_size(entries);
+ size_t i;
+
for (i = 0; i < vs; ++i) {
if (entries[i].ch == ch) {
t = &entries[i];
void correct_trie_set(correct_trie_t *t, const char *key, void * const value) {
const unsigned char *data = (const unsigned char*)key;
while (*data) {
- unsigned char ch = *data;
- correct_trie_t *entries = t->entries;
- const size_t vs = vec_size(t->entries);
- size_t i;
+ correct_trie_t *entries = t->entries;
+ const size_t vs = vec_size(entries);
+ unsigned char ch = *data;
+ size_t i;
+
for (i = 0; i < vs; ++i) {
if (entries[i].ch == ch) {
t = &entries[i];
}
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;
+ t = elem;
}
++data;
}
* 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) {
}
void correct_del(correct_trie_t* dictonary, size_t **data) {
- size_t i;
+ size_t i;
const size_t vs = vec_size(data);
+
for (i = 0; i < vs; i++)
mem_d(data[i]);
* because they're only valid after the first character is of a _, or
* alpha character.
*/
-static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_";
+static const char correct_alpha[] = "abcdefghijklmnopqrstuvwxyz"
+ "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+ "_"; /* TODO: Numbers ... */
/*
* correcting logic for the following forms of transformations:
* 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) {
- size_t itr;
- size_t len = strlen(ident);
+ size_t itr = 0;
+ const size_t len = strlen(ident);
- for (itr = 0; itr < len; itr++) {
+ for (; itr < len; itr++) {
char *a = (char*)correct_pool_alloc(len+1);
memcpy(a, ident, itr);
memcpy(a + itr, ident + itr + 1, len - itr);
}
static size_t correct_transposition(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t len = strlen(ident);
+ size_t itr = 0;
+ const size_t len = strlen(ident);
- for (itr = 0; itr < len - 1; itr++) {
+ for (; itr < len - 1; itr++) {
char tmp;
char *a = (char*)correct_pool_alloc(len+1);
memcpy(a, ident, len+1);
}
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);
+ size_t itr = 0;
+ size_t jtr = 0;
+ size_t ktr = 0;
+ const size_t len = strlen(ident);
- for (itr = 0, ktr = 0; itr < len; itr++) {
+ 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);
}
static size_t correct_insertion(const char *ident, char **array, size_t index) {
- size_t itr;
- size_t jtr;
- size_t ktr;
- const size_t len = strlen(ident);
+ size_t itr = 0;
+ size_t jtr = 0;
+ size_t ktr = 0;
+ const size_t len = strlen(ident);
- for (itr = 0, ktr = 0; itr <= len; itr++) {
+ 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);
*/
static int correct_exist(char **array, size_t rows, char *ident) {
size_t itr;
- for (itr = 0; itr < rows; itr++)
- if (!strcmp(array[itr], ident))
+ /*
+ * As an experiment I tried the following assembly for memcmp here:
+ *
+ * correct_cmp_loop:
+ * incl %eax ; eax = LHS
+ * incl %edx ; edx = LRS
+ * cmpl %eax, %ebx ; ebx = &LHS[END_POS]
+ *
+ * jbe correct_cmp_eq
+ * movb (%edx), %cl ; micro-optimized on even atoms :-)
+ * cmpb %cl, (%eax) ; ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+ * jg correct_cmp_gt
+ * jge correct_cmp_loop
+ * ...
+ *
+ * Despite how much optimization went in to this, the speed was the
+ * being conflicted by the strlen(ident) used for &LHS[END_POS]
+ * If we could eliminate the strlen with what I suggested on line
+ * 311 ... we can accelerate this whole damn thing quite a bit.
+ *
+ * However there is still something we can do here that does give
+ * us a little more speed. Although one more branch, we know for
+ * sure there is at least one byte to compare, if that one byte
+ * simply isn't the same we can skip the full check. Which means
+ * we skip a whole strlen call.
+ */
+ for (itr = 0; itr < rows; itr++) {
+ if (!memcmp(array[itr], ident, strlen(ident)))
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+1 < oldallocated)
return res;
- *allocated += 32;
- out = correct_pool_alloc(sizeof(*res) * *allocated);
+ out = 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) {
- size_t itr;
- size_t jtr;
- size_t len;
- size_t row;
+ 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 = 0, len = 0; itr < rows; itr++) {
+ for (; itr < rows; itr++) {
end = correct_edit(array[itr]);
row = correct_size(array[itr]);
+ /* removing jtr=0 here speeds it up by 100ms O_o */
for (jtr = 0; jtr < row; jtr++) {
if (correct_find(table, end[jtr]) && !correct_exist(res, len, end[jtr])) {
res = correct_known_resize(res, &nxt, len+1);
}
static char *correct_maximum(correct_trie_t* table, char **array, size_t rows) {
- char *str = NULL;
- size_t *itm = NULL;
- size_t itr;
- size_t top;
+ 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];
*
* the add function works the same. Except the identifier is used to
* add to the dictonary.
- */
-
+ */
char *correct_str(correct_trie_t* table, const char *ident) {
- char **e1;
- char **e2;
- char *e1ident;
- char *e2ident;
-
- size_t e1rows = 0;
- size_t e2rows = 0;
+ char **e1 = NULL;
+ char **e2 = NULL;
+ char *e1ident = NULL;
+ char *e2ident = NULL;
+ size_t e1rows = 0;
+ size_t e2rows = 0;
correct_pool_new();