* 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.
*
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[] = "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
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) {
- 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];
- ++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) {
- 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];
- 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;
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
* 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) {
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 next;
char **find = (char**)correct_pool_alloc(correct_size(ident) * sizeof(char*));
if (!find)
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);
return find;
}
* 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, size_t rows, char *ident) {
size_t itr;
/*
* As an experiment I tried the following assembly for memcmp here:
* jge correct_cmp_loop
* ...
*
- * Despite how much optimization went in to this, the speed was the
+ * Despite how much optimization went in to this, the speed was
* 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.
if (size < oldallocated)
return res;
- out = correct_pool_alloc(sizeof(*res) * oldallocated + 32);
+ 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 **res = (char**)correct_pool_alloc(sizeof(char *) * nxt);
char **end = 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];
+ else {
+ end = correct_edit(array[itr]);
+ vec_push(corr->edits, end);
+ }
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);
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;
* takes a table for the dictonary a vector of sizes (used for internal
* 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;
+}
+
+void correct_free(correction_t *c)
+{
+ vec_free(c->edits);
+ 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;
size_t e1rows = 0;
size_t e2rows = 0;
- correct_pool_new();
-
/* 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);
+ vec_push(corr->edits, e1);
+ }
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);
}