4 * Copyright 2004 Develer S.r.l. (http://www.develer.com/)
5 * Copyright 2004 Giovanni Bajo
9 * \brief Portable hash table implementation
11 * Some rationales of our choices in implementation:
13 * \li For embedded systems, it is vital to allocate the table in static memory. To do
14 * so, it is necessary to expose the \c HashNode and \c HashTable structures in the header file.
15 * Nevertheless, they should be used as opaque types (that is, the users should not
16 * access the structure fields directly).
18 * \li To statically allocate the structures, a macro is provided. With this macro, we
19 * are hiding completely \c HashNode to the user (who only manipulates \c HashTable). Without
20 * the macro, the user would have had to define both the \c HashNode and the \c HashTable
21 * manually, and pass both of them to \c ht_init() (which would have created the link between
22 * the two). Instead, the link is created with a literal initialization.
24 * \li The hash table is created as power of two to remove the divisions from the code.
25 * Of course, hash functions work at their best when the table size is a prime number.
26 * When calculating the modulus to convert the hash value to an index, the actual operation
27 * becomes a bitwise AND: this is fast, but truncates the value losing bits. Thus, the higher
28 * bits are first "merged" with the lower bits through some XOR operations (see the last line of
31 * \li To minimize the memory occupation, there is no flag to set for the empty node. An
32 * empty node is recognized by its data pointer set to NULL. It is then invalid to store
33 * NULL as data pointer in the table.
35 * \li The visiting interface through iterators is implemented with pass-by-value semantic.
36 * While this is overkill for medium-to-stupid compilers, it is the best designed from an
37 * user point of view. Moreover, being totally inlined (defined completely in the header),
38 * even a stupid compiler should be able to perform basic optimizations on it.
39 * We thought about using a pass-by-pointer semantic but it was much more awful to use, and
40 * the compiler is then forced to spill everything to the stack (unless it is *very* smart).
42 * \li The current implementation allows to either store the key internally (that is, copy
43 * the key within the hash table) or keep it external (that is, a hook is used to extract
44 * the key from the data in the node). The former is more memory-hungry of course, as it
45 * allocated static space to store the key copies. The overhead to keep both methods at
46 * the same time is minimal:
48 * <li>There is a run-time check in node_get_key which is execute per each node visited.</li>
49 * <li>Theoretically, there is no memory overhead. In practice, there were no
50 * flags in \c struct HashTable till now, so we had to add a first bit flag, but the
51 * overhead will disappear if a second flag is added for a different reason later.</li>
52 * <li>There is a little interface overhead, since we have two different versions of
53 * \c ht_insert(), one with the key passed as parameter and one without, but in
54 * the common case (external keys) both can be used.</li>
59 * \author Giovanni Bajo <rasky@develer.com>
64 *#* Revision 1.6 2006/06/01 12:27:39 marco
65 *#* Added utilities for protocols
69 #include "hashtable.h"
70 #include <cfg/debug.h>
71 #include <cfg/compiler.h>
76 #define ROTATE_LEFT_16(num, count) (((num) << (count)) | ((num) >> (16-(count))))
77 #define ROTATE_RIGHT_16(num, count) ROTATE_LEFT_16(num, 16-(count))
79 typedef const void** HashNodePtr;
80 #define NODE_EMPTY(node) (!*(node))
81 #define HT_HAS_INTERNAL_KEY(ht) (CONFIG_HT_OPTIONAL_INTERNAL_KEY && ht->flags.key_internal)
83 /*! For hash tables with internal keys, compute the pointer to the internal key for a given \a node. */
84 INLINE uint8_t *key_internal_get_ptr(struct HashTable *ht, HashNodePtr node)
86 uint8_t* key_buf = ht->key_data.mem;
89 // Compute the index of the node and use it to move within the whole key buffer
90 index = node - &ht->mem[0];
91 ASSERT(index < (size_t)(1 << ht->max_elts_log2));
92 key_buf += index * (INTERNAL_KEY_MAX_LENGTH + 1);
98 INLINE void node_get_key(struct HashTable* ht, HashNodePtr node, const void** key, uint8_t* key_length)
100 if (HT_HAS_INTERNAL_KEY(ht))
102 uint8_t* k = key_internal_get_ptr(ht, node);
104 // Key has its length stored in the first byte
109 *key = ht->key_data.hook(*node, key_length);
113 INLINE bool node_key_match(struct HashTable* ht, HashNodePtr node, const void* key, uint8_t key_length)
118 node_get_key(ht, node, &key2, &key2_length);
120 return (key_length == key2_length && memcmp(key, key2, key_length) == 0);
124 static uint16_t calc_hash(const void* _key, uint8_t key_length)
126 const char* key = (const char*)_key;
127 uint16_t hash = key_length;
129 int len = (int)key_length;
131 for (i = 0; i < len; ++i)
132 hash = ROTATE_LEFT_16(hash, 4) ^ key[i];
134 return hash ^ (hash >> 6) ^ (hash >> 13);
138 static HashNodePtr perform_lookup(struct HashTable* ht,
139 const void* key, uint8_t key_length)
141 uint16_t hash = calc_hash(key, key_length);
142 uint16_t mask = ((1 << ht->max_elts_log2) - 1);
143 uint16_t index = hash & mask;
144 uint16_t first_index = index;
148 // Fast-path optimization: we check immediately if the current node
149 // is the one we were looking for, so we save the computation of the
150 // increment step in the common case.
151 node = &ht->mem[index];
153 || node_key_match(ht, node, key, key_length))
156 // Increment while going through the hash table in case of collision.
157 // This implements the double-hash technique: we use the higher part
158 // of the hash as a step increment instead of just going to the next
159 // element, to minimize the collisions.
160 // Notice that the number must be odd to be sure that the whole table
161 // is traversed. Actually MCD(table_size, step) must be 1, but
162 // table_size is always a power of 2, so we just ensure that step is
163 // never a multiple of 2.
164 step = (ROTATE_RIGHT_16(hash, ht->max_elts_log2) & mask) | 1;
171 node = &ht->mem[index];
173 || node_key_match(ht, node, key, key_length))
176 // The check is done after the key compare. This actually causes
177 // one more compare in the case the table is full (since the first
178 // element was compared at the very start, and then at the end),
179 // but it makes faster the common path where we enter this loop
180 // for the first time, and index will not match first_index for
182 } while (index != first_index);
188 void ht_init(struct HashTable* ht)
190 memset(ht->mem, 0, sizeof(ht->mem[0]) * (1 << ht->max_elts_log2));
194 static bool insert(struct HashTable* ht, const void* key, uint8_t key_length, const void* data)
201 if (HT_HAS_INTERNAL_KEY(ht))
202 key_length = MIN(key_length, (uint8_t)INTERNAL_KEY_MAX_LENGTH);
204 node = perform_lookup(ht, key, key_length);
208 if (HT_HAS_INTERNAL_KEY(ht))
210 uint8_t* k = key_internal_get_ptr(ht, node);
212 memcpy(k, key, key_length);
220 bool ht_insert_with_key(struct HashTable* ht, const void* key, uint8_t key_length, const void* data)
223 if (!HT_HAS_INTERNAL_KEY(ht))
225 // Construct a fake node and use it to match the key
226 HashNodePtr node = &data;
227 if (!node_key_match(ht, node, key, key_length))
229 ASSERT2(0, "parameter key is different from the external key");
235 return insert(ht, key, key_length, data);
239 bool ht_insert(struct HashTable* ht, const void* data)
245 if (HT_HAS_INTERNAL_KEY(ht))
247 ASSERT("parameter cannot be a hash table with internal keys - use ht_insert_with_key()"
253 key = ht->key_data.hook(data, &key_length);
255 return insert(ht, key, key_length, data);
259 const void* ht_find(struct HashTable* ht, const void* key, uint8_t key_length)
263 if (HT_HAS_INTERNAL_KEY(ht))
264 key_length = MIN(key_length, (uint8_t)INTERNAL_KEY_MAX_LENGTH);
266 node = perform_lookup(ht, key, key_length);
268 if (!node || NODE_EMPTY(node))
281 static const void* test_get_key(const void* ptr, uint8_t* length)
288 #define NUM_ELEMENTS 256
289 DECLARE_HASHTABLE_STATIC(test1, 256, test_get_key);
290 DECLARE_HASHTABLE_INTERNALKEY_STATIC(test2, 256);
292 static char data[NUM_ELEMENTS][10];
293 static char keydomain[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
295 static bool single_test(void)
302 for (i=0;i<NUM_ELEMENTS;i++)
309 klen = (rand() % 8) + 1;
311 data[i][k] = keydomain[rand() % (sizeof(keydomain)-1)];
313 } while (ht_find_str(&test1, data[i]) != NULL);
315 ASSERT(ht_insert(&test1, data[i]));
316 ASSERT(ht_insert_str(&test2, data[i], data[i]));
319 for (i=0;i<NUM_ELEMENTS;i++)
321 char *found1, *found2;
323 found1 = (char*)ht_find_str(&test1, data[i]);
324 if (strcmp(found1, data[i]) != 0)
326 ASSERT(strcmp(found1,data[i]) == 0);
330 found2 = (char*)ht_find_str(&test2, data[i]);
331 if (strcmp(found2, data[i]) != 0)
333 ASSERT(strcmp(found2,data[i]) == 0);
341 static uint16_t rand_seeds[] = { 1, 42, 666, 0xDEAD, 0xBEEF, 0x1337, 0xB00B };
347 for (i=0;i<countof(rand_seeds);++i)
349 srand(rand_seeds[i]);
352 kprintf("ht_test failed\n");
357 kprintf("ht_test successful\n");