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 1.8 2007/02/06 16:05:01 asterix
65 *#* Replaced ROTATE_* with ROT* defined in macros.h
67 *#* Revision 1.7 2006/07/19 12:56:27 bernie
68 *#* Convert to new Doxygen style.
70 *#* Revision 1.6 2006/06/01 12:27:39 marco
71 *#* Added utilities for protocols
75 #include "hashtable.h"
76 #include <cfg/debug.h>
77 #include <cfg/compiler.h>
78 #include <cfg/macros.h> //ROTL(), ROTR();
84 typedef const void** HashNodePtr;
85 #define NODE_EMPTY(node) (!*(node))
86 #define HT_HAS_INTERNAL_KEY(ht) (CONFIG_HT_OPTIONAL_INTERNAL_KEY && ht->flags.key_internal)
88 /** For hash tables with internal keys, compute the pointer to the internal key for a given \a node. */
89 INLINE uint8_t *key_internal_get_ptr(struct HashTable *ht, HashNodePtr node)
91 uint8_t* key_buf = ht->key_data.mem;
94 // Compute the index of the node and use it to move within the whole key buffer
95 index = node - &ht->mem[0];
96 ASSERT(index < (size_t)(1 << ht->max_elts_log2));
97 key_buf += index * (INTERNAL_KEY_MAX_LENGTH + 1);
103 INLINE void node_get_key(struct HashTable* ht, HashNodePtr node, const void** key, uint8_t* key_length)
105 if (HT_HAS_INTERNAL_KEY(ht))
107 uint8_t* k = key_internal_get_ptr(ht, node);
109 // Key has its length stored in the first byte
114 *key = ht->key_data.hook(*node, key_length);
118 INLINE bool node_key_match(struct HashTable* ht, HashNodePtr node, const void* key, uint8_t key_length)
123 node_get_key(ht, node, &key2, &key2_length);
125 return (key_length == key2_length && memcmp(key, key2, key_length) == 0);
129 static uint16_t calc_hash(const void* _key, uint8_t key_length)
131 const char* key = (const char*)_key;
132 uint16_t hash = key_length;
134 int len = (int)key_length;
136 for (i = 0; i < len; ++i)
137 hash = ROTL(hash, 4) ^ key[i];
139 return hash ^ (hash >> 6) ^ (hash >> 13);
143 static HashNodePtr perform_lookup(struct HashTable* ht,
144 const void* key, uint8_t key_length)
146 uint16_t hash = calc_hash(key, key_length);
147 uint16_t mask = ((1 << ht->max_elts_log2) - 1);
148 uint16_t index = hash & mask;
149 uint16_t first_index = index;
153 // Fast-path optimization: we check immediately if the current node
154 // is the one we were looking for, so we save the computation of the
155 // increment step in the common case.
156 node = &ht->mem[index];
158 || node_key_match(ht, node, key, key_length))
161 // Increment while going through the hash table in case of collision.
162 // This implements the double-hash technique: we use the higher part
163 // of the hash as a step increment instead of just going to the next
164 // element, to minimize the collisions.
165 // Notice that the number must be odd to be sure that the whole table
166 // is traversed. Actually MCD(table_size, step) must be 1, but
167 // table_size is always a power of 2, so we just ensure that step is
168 // never a multiple of 2.
169 step = (ROTR(hash, ht->max_elts_log2) & mask) | 1;
176 node = &ht->mem[index];
178 || node_key_match(ht, node, key, key_length))
181 // The check is done after the key compare. This actually causes
182 // one more compare in the case the table is full (since the first
183 // element was compared at the very start, and then at the end),
184 // but it makes faster the common path where we enter this loop
185 // for the first time, and index will not match first_index for
187 } while (index != first_index);
193 void ht_init(struct HashTable* ht)
195 memset(ht->mem, 0, sizeof(ht->mem[0]) * (1 << ht->max_elts_log2));
199 static bool insert(struct HashTable* ht, const void* key, uint8_t key_length, const void* data)
206 if (HT_HAS_INTERNAL_KEY(ht))
207 key_length = MIN(key_length, (uint8_t)INTERNAL_KEY_MAX_LENGTH);
209 node = perform_lookup(ht, key, key_length);
213 if (HT_HAS_INTERNAL_KEY(ht))
215 uint8_t* k = key_internal_get_ptr(ht, node);
217 memcpy(k, key, key_length);
225 bool ht_insert_with_key(struct HashTable* ht, const void* key, uint8_t key_length, const void* data)
228 if (!HT_HAS_INTERNAL_KEY(ht))
230 // Construct a fake node and use it to match the key
231 HashNodePtr node = &data;
232 if (!node_key_match(ht, node, key, key_length))
234 ASSERT2(0, "parameter key is different from the external key");
240 return insert(ht, key, key_length, data);
244 bool ht_insert(struct HashTable* ht, const void* data)
250 if (HT_HAS_INTERNAL_KEY(ht))
252 ASSERT("parameter cannot be a hash table with internal keys - use ht_insert_with_key()"
258 key = ht->key_data.hook(data, &key_length);
260 return insert(ht, key, key_length, data);
264 const void* ht_find(struct HashTable* ht, const void* key, uint8_t key_length)
268 if (HT_HAS_INTERNAL_KEY(ht))
269 key_length = MIN(key_length, (uint8_t)INTERNAL_KEY_MAX_LENGTH);
271 node = perform_lookup(ht, key, key_length);
273 if (!node || NODE_EMPTY(node))
286 static const void* test_get_key(const void* ptr, uint8_t* length)
293 #define NUM_ELEMENTS 256
294 DECLARE_HASHTABLE_STATIC(test1, 256, test_get_key);
295 DECLARE_HASHTABLE_INTERNALKEY_STATIC(test2, 256);
297 static char data[NUM_ELEMENTS][10];
298 static char keydomain[] = "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789";
300 static bool single_test(void)
307 for (i=0;i<NUM_ELEMENTS;i++)
314 klen = (rand() % 8) + 1;
316 data[i][k] = keydomain[rand() % (sizeof(keydomain)-1)];
318 } while (ht_find_str(&test1, data[i]) != NULL);
320 ASSERT(ht_insert(&test1, data[i]));
321 ASSERT(ht_insert_str(&test2, data[i], data[i]));
324 for (i=0;i<NUM_ELEMENTS;i++)
326 char *found1, *found2;
328 found1 = (char*)ht_find_str(&test1, data[i]);
329 if (strcmp(found1, data[i]) != 0)
331 ASSERT(strcmp(found1,data[i]) == 0);
335 found2 = (char*)ht_find_str(&test2, data[i]);
336 if (strcmp(found2, data[i]) != 0)
338 ASSERT(strcmp(found2,data[i]) == 0);
346 static uint16_t rand_seeds[] = { 1, 42, 666, 0xDEAD, 0xBEEF, 0x1337, 0xB00B };
352 for (i=0;i<countof(rand_seeds);++i)
354 srand(rand_seeds[i]);
357 kprintf("ht_test failed\n");
362 kprintf("ht_test successful\n");