/* INTERFACE:
- Hash tables are an implementation technique used to implement
- mapping between objects. Assuming a good hashing function is used,
- they provide near-constant-time access and storing of information.
- Duplicate keys are not allowed.
-
- This file defines the following entry points: hash_table_new
- creates a hash table, and hash_table_destroy deletes it.
- hash_table_put establishes a mapping between a key and a value.
- hash_table_get retrieves the value that corresponds to a key.
- hash_table_contains queries whether a key is stored in a table at
- all. hash_table_remove removes a mapping that corresponds to a
- key. hash_table_map allows you to map through all the entries in a
- hash table. hash_table_clear clears all the entries from the hash
- table.
-
- The number of mappings in a table is not limited, except by the
- amount of memory. As you add new elements to a table, it regrows
- as necessary. If you have an idea about how many elements you will
- store, you can provide a hint to hash_table_new().
-
- The hashing and equality functions depend on the type of key and
- are normally provided by the user. For the special (and frequent)
- case of using string keys, you can use the pre-canned
- make_string_hash_table(), which provides an efficient string
- hashing function, and a string equality wrapper around strcmp().
-
- When specifying your hash and test functions, make sure the
- following holds true:
-
- - The test function returns non-zero for keys that are considered
- "equal", zero otherwise.
-
- - The hash function returns a number that represents the
- "distinctness" of the object. In more precise terms, it means
- that for any two objects that test "equal" under the test
- function, the hash function MUST produce the same result.
-
- This does not mean that each distinct object must produce a
- distinct value, only that non-distinct objects must produce the
- same values! For instance, a hash function that returns 0 for
- any given object is a perfectly valid (albeit extremely bad) hash
- function. A hash function that hashes a string by adding up all
- its characters is another example of a valid (but quite bad) hash
- function.
-
- The above stated rule is quite easy to enforce. For example, if
- your testing function compares strings case-insensitively, all
- your function needs to do is lower-case the string characters
- before calculating a hash. That way you have easily guaranteed
- that case differences will not result in a different hash.
-
- - If you care about performance, choose a hash function with as
- good "spreading" as possible. A good hash function will react to
- even a small change in its input with a completely different
- resulting hash. Finally, don't make the hash function itself
- overly slow, because you'll be incurring a non-negligible
- overhead to reads and writes to the hash table.
+ Hash tables are a technique used to implement mapping between
+ objects with near-constant-time access and storage. The table
+ associates keys to values, and a value can be very quickly
+ retrieved by providing the key. Fast lookup tables are typically
+ implemented as hash tables.
+
+ The entry points are
+ hash_table_new -- creates the table.
+ hash_table_destroy -- destroys the table.
+ hash_table_put -- establishes or updates key->value mapping.
+ hash_table_get -- retrieves value of key.
+ hash_table_get_pair -- get key/value pair for key.
+ hash_table_contains -- test whether the table contains key.
+ hash_table_remove -- remove the key->value mapping for key.
+ hash_table_map -- iterate through table mappings.
+ hash_table_clear -- clear hash table contents.
+ hash_table_count -- return the number of entries in the table.
+
+ The hash table grows internally as new entries are added and is not
+ limited in size, except by available memory. The table doubles
+ with each resize, which ensures that the amortized time per
+ operation remains constant.
+
+ By default, tables created by hash_table_new consider the keys to
+ be equal if their pointer values are the same. You can use
+ make_string_hash_table to create tables whose keys are considered
+ equal if their string contents are the same. In the general case,
+ the criterion of equality used to compare keys is specified at
+ table creation time with two callback functions, "hash" and "test".
+ The hash function transforms the key into an arbitrary number that
+ must be the same for two equal keys. The test function accepts two
+ keys and returns non-zero if they are to be considered equal.
Note that neither keys nor values are copied when inserted into the
hash table, so they must exist for the lifetime of the table. This
/* IMPLEMENTATION:
- All the hash mappings (key-value pairs of pointers) are stored in a
- contiguous array. The position of each mapping is determined by
- the hash value of its key and the size of the table: location :=
- hash(key) % size. If two different keys end up on the same
- position (hash collision), the one that came second is placed at
- the next empty position following the occupied place. This
- collision resolution technique is called "linear probing".
+ The hash table is implemented as an open-addressed table with
+ linear probing collision resolution.
- There are more advanced collision resolution mechanisms (quadratic
+ For those not up to CS parlance, it means that all the hash entries
+ (pairs of pointers key and value) are stored in a contiguous array.
+ The position of each mapping is determined by the hash value of its
+ key and the size of the table: location := hash(key) % size. If
+ two different keys end up on the same position (collide), the one
+ that came second is placed at the next empty position following the
+ occupied place. This collision resolution technique is called
+ "linear probing".
+
+ There are more advanced collision resolution methods (quadratic
probing, double hashing), but we don't use them because they incur
- more non-sequential access to the array, which results in worse
+ more non-sequential access to the array, which results in worse CPU
cache behavior. Linear probing works well as long as the
- fullness/size ratio is kept below 75%. We make sure to regrow or
- rehash the hash table whenever this threshold is exceeded.
-
- Collisions make deletion tricky because finding collisions again
- relies on new empty spots not being created. That's why
- hash_table_remove is careful to rehash the mappings that follow the
- deleted one. */
+ count/size ratio (fullness) is kept below 75%. We make sure to
+ grow and rehash the table whenever this threshold is exceeded.
+
+ Collisions make deletion tricky because clearing a position
+ followed by a colliding entry would make the position seem empty
+ and the colliding entry not found. One solution is to leave a
+ "tombstone" instead of clearing the entry, and another is to
+ carefully rehash the entries immediately following the deleted one.
+ We use the latter method because it results in less bookkeeping and
+ faster retrieval at the (slight) expense of deletion. */
+
+/* Maximum allowed fullness: when hash table's fullness exceeds this
+ value, the table is resized. */
+#define HASH_MAX_FULLNESS 0.75
+
+/* The hash table size is multiplied by this factor (and then rounded
+ to the next prime) with each resize. This guarantees infrequent
+ resizes. */
+#define HASH_RESIZE_FACTOR 2
struct mapping {
void *key;
unsigned long (*hash_function) PARAMS ((const void *));
int (*test_function) PARAMS ((const void *, const void *));
- int size; /* size of the array */
- int count; /* number of non-empty, non-deleted
- fields. */
+ int size; /* size of the array. */
+ int count; /* number of non-empty entries. */
int resize_threshold; /* after size exceeds this number of
entries, resize the table. */
struct mapping *mappings; /* the array of mapping pairs. */
};
-#define EMPTY_MAPPING_P(mp) ((mp)->key == NULL)
-#define NEXT_MAPPING(mp, mappings, size) (mp == mappings + (size - 1) \
- ? mappings : mp + 1)
+/* We use all-bit-set marker to mean that a mapping is empty. It is
+ (hopefully) illegal as a pointer, and it allows the users to use
+ NULL (as well as any non-negative integer) as key. */
+#define NON_EMPTY(mp) (mp->key != (void *)~(unsigned long)0)
+/* "Next" mapping is the mapping after MP, but wrapping back to
+ MAPPINGS when MP would reach MAPPINGS+SIZE. */
+#define NEXT_MAPPING(mp, mappings, size) (mp != mappings + (size - 1) \
+ ? mp + 1 : mappings)
+
+/* Loop over non-empty mappings starting at MP. */
#define LOOP_NON_EMPTY(mp, mappings, size) \
- for (; !EMPTY_MAPPING_P (mp); mp = NEXT_MAPPING (mp, mappings, size))
+ for (; NON_EMPTY (mp); mp = NEXT_MAPPING (mp, mappings, size))
-/* #### We might want to multiply with the "golden ratio" here to get
- better randomness for keys that do not result from a good hash
- function. This is currently not a problem in Wget because we only
- use the string hash tables. */
+/* #### Some implementations multiply the hash with the "golden ratio"
+ of the table to get better spread for keys that do not come from a
+ good hashing source. I'm not sure if that is necessary for the
+ hash functions we use. */
#define HASH_POSITION(ht, key) (ht->hash_function (key) % ht->size)
table does not contain all primes in range, just a selection useful
for this purpose.
- PRIME_OFFSET is a micro-optimization: if specified, it starts the
+ PRIME_OFFSET is a minor optimization: if specified, it starts the
search for the prime number beginning with the specific offset in
the prime number table. The final offset is stored in the same
variable. */
return 0;
}
-/* Create a hash table of INITIAL_SIZE with hash function
- HASH_FUNCTION and test function TEST_FUNCTION. INITIAL_SIZE will
- be rounded to the next prime, so you don't have to worry about it
- being a prime number.
+static unsigned long ptrhash PARAMS ((const void *));
+static int ptrcmp PARAMS ((const void *, const void *));
+
+/* Create a hash table with hash function HASH_FUNCTION and test
+ function TEST_FUNCTION. The table is empty (its count is 0), but
+ pre-allocated to store at least ITEMS items.
- Consequently, if you wish to start out with a "small" table which
- will be regrown as needed, specify INITIAL_SIZE 0.
+ ITEMS is the number of items that the table can accept without
+ needing to resize. It is useful when creating a table that is to
+ be immediately filled with a known number of items. In that case,
+ the regrows are a waste of time, and specifying ITEMS correctly
+ will avoid them altogether.
+
+ Note that hash tables grow dynamically regardless of ITEMS. The
+ only use of ITEMS is to preallocate the table and avoid unnecessary
+ dynamic regrows. Don't bother making ITEMS prime because it's not
+ used as size unchanged. To start with a small table that grows as
+ needed, simply specify zero ITEMS.
If HASH_FUNCTION is not provided, identity table is assumed,
i.e. key pointers are compared as keys. If you want strings with
equal contents to hash the same, use make_string_hash_table. */
struct hash_table *
-hash_table_new (int initial_size,
+hash_table_new (int items,
unsigned long (*hash_function) (const void *),
int (*test_function) (const void *, const void *))
{
- struct hash_table *ht
- = (struct hash_table *)xmalloc (sizeof (struct hash_table));
+ int size;
+ struct hash_table *ht = xnew (struct hash_table);
ht->hash_function = hash_function ? hash_function : ptrhash;
ht->test_function = test_function ? test_function : ptrcmp;
+ /* If the size of struct hash_table ever becomes a concern, this
+ field can go. (Wget doesn't create many hashes.) */
ht->prime_offset = 0;
- ht->size = prime_size (initial_size, &ht->prime_offset);
- ht->resize_threshold = ht->size * 3 / 4;
- ht->count = 0;
+ /* Calculate the size that ensures that the table will store at
+ least ITEMS keys without the need to resize. */
+ size = 1 + items / HASH_MAX_FULLNESS;
+ size = prime_size (size, &ht->prime_offset);
+ ht->size = size;
+ ht->resize_threshold = size * HASH_MAX_FULLNESS;
+ /*assert (ht->resize_threshold >= items);*/
- ht->mappings = xmalloc (ht->size * sizeof (struct mapping));
- memset (ht->mappings, '\0', ht->size * sizeof (struct mapping));
+ ht->mappings = xnew_array (struct mapping, ht->size);
+ /* Mark mappings as empty. We use 0xff rather than 0 to mark empty
+ keys because it allows us to store NULL keys to the table. */
+ memset (ht->mappings, 255, size * sizeof (struct mapping));
+
+ ht->count = 0;
return ht;
}
xfree (ht);
}
-/* The heart of almost all functions in this file -- find the mapping
- whose KEY is equal to key, using linear probing. Returns the
- mapping that matches KEY, or NULL if none matches. */
+/* The heart of most functions in this file -- find the mapping whose
+ KEY is equal to key, using linear probing. Returns the mapping
+ that matches KEY, or the first empty mapping if none matches. */
static inline struct mapping *
find_mapping (const struct hash_table *ht, const void *key)
LOOP_NON_EMPTY (mp, mappings, size)
if (equals (key, mp->key))
- return mp;
- return NULL;
+ break;
+ return mp;
}
/* Get the value that corresponds to the key KEY in the hash table HT.
hash_table_get (const struct hash_table *ht, const void *key)
{
struct mapping *mp = find_mapping (ht, key);
- if (mp)
+ if (NON_EMPTY (mp))
return mp->value;
else
return NULL;
void *orig_key, void *value)
{
struct mapping *mp = find_mapping (ht, lookup_key);
-
- if (mp)
+ if (NON_EMPTY (mp))
{
if (orig_key)
*(void **)orig_key = mp->key;
int
hash_table_contains (const struct hash_table *ht, const void *key)
{
- return find_mapping (ht, key) != NULL;
+ struct mapping *mp = find_mapping (ht, key);
+ return NON_EMPTY (mp);
}
/* Grow hash table HT as necessary, and rehash all the key-value
struct mapping *mp, *mappings;
int newsize;
- newsize = prime_size (ht->size * 2, &ht->prime_offset);
+ newsize = prime_size (ht->size * HASH_RESIZE_FACTOR, &ht->prime_offset);
#if 0
printf ("growing from %d to %d; fullness %.2f%% to %.2f%%\n",
ht->size, newsize,
- (double)100 * ht->count / ht->size,
- (double)100 * ht->count / newsize);
+ 100.0 * ht->count / ht->size,
+ 100.0 * ht->count / newsize);
#endif
ht->size = newsize;
- ht->resize_threshold = newsize * 3 / 4;
+ ht->resize_threshold = newsize * HASH_MAX_FULLNESS;
- mappings = xmalloc (ht->size * sizeof (struct mapping));
- memset (mappings, '\0', ht->size * sizeof (struct mapping));
+ mappings = xnew_array (struct mapping, newsize);
+ memset (mappings, 255, newsize * sizeof (struct mapping));
ht->mappings = mappings;
for (mp = old_mappings; mp < old_end; mp++)
- if (!EMPTY_MAPPING_P (mp))
+ if (NON_EMPTY (mp))
{
struct mapping *new_mp = mappings + HASH_POSITION (ht, mp->key);
- /* We don't need to call test function and worry about
- collisions because all the keys come from the hash table
- and are therefore guaranteed to be unique. */
+ /* We don't need to test for uniqueness of keys because they
+ come from the hash table and are therefore known to be
+ unique. */
LOOP_NON_EMPTY (new_mp, mappings, newsize)
;
*new_mp = *mp;
void
hash_table_put (struct hash_table *ht, const void *key, void *value)
{
- struct mapping *mappings = ht->mappings;
- int size = ht->size;
- int (*equals) PARAMS ((const void *, const void *)) = ht->test_function;
-
- struct mapping *mp = mappings + HASH_POSITION (ht, key);
+ struct mapping *mp = find_mapping (ht, key);
+ if (NON_EMPTY (mp))
+ {
+ /* update existing item */
+ mp->key = (void *)key; /* const? */
+ mp->value = value;
+ return;
+ }
- LOOP_NON_EMPTY (mp, mappings, size)
- if (equals (key, mp->key))
- {
- mp->key = (void *)key; /* const? */
- mp->value = value;
- return;
- }
+ /* If adding the item would make the table exceed max. fullness,
+ grow the table first. */
+ if (ht->count >= ht->resize_threshold)
+ {
+ grow_hash_table (ht);
+ mp = find_mapping (ht, key);
+ }
+ /* add new item */
++ht->count;
mp->key = (void *)key; /* const? */
mp->value = value;
-
- if (ht->count > ht->resize_threshold)
- /* When table is 75% full, regrow it. */
- grow_hash_table (ht);
}
/* Remove a mapping that matches KEY from HT. Return 0 if there was
hash_table_remove (struct hash_table *ht, const void *key)
{
struct mapping *mp = find_mapping (ht, key);
- if (!mp)
+ if (!NON_EMPTY (mp))
return 0;
else
{
}
/* Map MAPFUN over all the mappings in hash table HT. MAPFUN is
- called with three arguments: the key, the value, and the CLOSURE.
+ called with three arguments: the key, the value, and MAPARG.
It is undefined what happens if you add or remove entries in the
hash table while hash_table_map is running. The exception is the
void
hash_table_map (struct hash_table *ht,
int (*mapfun) (void *, void *, void *),
- void *closure)
+ void *maparg)
{
struct mapping *mp = ht->mappings;
struct mapping *end = ht->mappings + ht->size;
for (; mp < end; mp++)
- if (!EMPTY_MAPPING_P (mp))
+ if (NON_EMPTY (mp))
{
void *key;
repeat:
key = mp->key;
- if (mapfun (key, mp->value, closure))
+ if (mapfun (key, mp->value, maparg))
return;
/* hash_table_remove might have moved the adjacent
mappings. */
- if (mp->key != key && !EMPTY_MAPPING_P (mp))
+ if (mp->key != key && NON_EMPTY (mp))
goto repeat;
}
}
don't strictly belong to this file. However, this is as good a
place for them as any. */
+/* Rules for creating custom hash and test functions:
+
+ - The test function returns non-zero for keys that are considered
+ "equal", zero otherwise.
+
+ - The hash function returns a number that represents the
+ "distinctness" of the object. In more precise terms, it means
+ that for any two objects that test "equal" under the test
+ function, the hash function MUST produce the same result.
+
+ This does not mean that all different objects must produce
+ different values (that would be "perfect" hashing), only that
+ non-distinct objects must produce the same values! For instance,
+ a hash function that returns 0 for any given object is a
+ perfectly valid (albeit extremely bad) hash function. A hash
+ function that hashes a string by adding up all its characters is
+ another example of a valid (but quite bad) hash function.
+
+ It is not hard to make hash and test functions agree about
+ equality. For example, if the test function compares strings
+ case-insensitively, the hash function can lower-case the
+ characters when calculating the hash value. That ensures that
+ two strings differing only in case will hash the same.
+
+ - If you care about performance, choose a hash function with as
+ good "spreading" as possible. A good hash function will use all
+ the bits of the input when calculating the hash, and will react
+ to even small changes in input with a completely different
+ output. Finally, don't make the hash function itself overly
+ slow, because you'll be incurring a non-negligible overhead to
+ all hash table operations. */
+
/*
* Support for hash tables whose keys are strings.
*
return !strcmp ((const char *)s1, (const char *)s2);
}
-/* Return a hash table of initial size INITIAL_SIZE suitable to use
- strings as keys. */
+/* Return a hash table of preallocated to store at least ITEMS items
+ suitable to use strings as keys. */
struct hash_table *
-make_string_hash_table (int initial_size)
+make_string_hash_table (int items)
{
- return hash_table_new (initial_size, string_hash, string_cmp);
+ return hash_table_new (items, string_hash, string_cmp);
}
/*
string_cmp_nocase. */
struct hash_table *
-make_nocase_string_hash_table (int initial_size)
+make_nocase_string_hash_table (int items)
{
- return hash_table_new (initial_size, string_hash_nocase, string_cmp_nocase);
+ return hash_table_new (items, string_hash_nocase, string_cmp_nocase);
}
/* Hashing of pointers. Used for hash tables that are keyed by
pointer identity. (Common Lisp calls them EQ hash tables, and Java
calls them IdentityHashMaps.) */
-unsigned long
+static unsigned long
ptrhash (const void *ptr)
{
unsigned long key = (unsigned long)ptr;
return key;
}
-int
+static int
ptrcmp (const void *ptr1, const void *ptr2)
{
return ptr1 == ptr2;
}
-
-#if 0
-/* Currently unused: hashing of integers. */
-
-unsigned long
-inthash (unsigned int key)
-{
- key += (key << 12);
- key ^= (key >> 22);
- key += (key << 4);
- key ^= (key >> 9);
- key += (key << 10);
- key ^= (key >> 2);
- key += (key << 7);
- key ^= (key >> 12);
- return key;
-}
-#endif
\f
#ifdef STANDALONE
projects / wget / blobdiff