fixing hash function to do a better job on long lists.

This commit is contained in:
JeffBezanson 2009-05-20 04:30:00 +00:00
parent bfbbf051c9
commit ff650e3049
2 changed files with 43 additions and 37 deletions

View File

@ -259,9 +259,10 @@ value_t equal(value_t a, value_t b)
#define doublehash(a) int64to32hash(a)
#endif
// *ut means we had to start using the table
static uptrint_t bounded_hash(value_t a, int bound, int *ut)
// *oob: output argument, means we hit the limit specified by 'bound'
static uptrint_t bounded_hash(value_t a, int bound, int *oob)
{
*oob = 0;
double d;
numerictype_t nt;
size_t i, len;
@ -269,12 +270,7 @@ static uptrint_t bounded_hash(value_t a, int bound, int *ut)
cprim_t *cp;
void *data;
uptrint_t h = 0;
if (*ut) {
h = (uptrint_t)ptrhash_get(&equal_eq_hashtable, (void*)a);
if (h != (uptrint_t)HT_NOTFOUND)
return h;
}
int tg = tag(a);
int oob2, tg = tag(a);
switch(tg) {
case TAG_NUM :
case TAG_NUM1:
@ -282,7 +278,7 @@ static uptrint_t bounded_hash(value_t a, int bound, int *ut)
return doublehash(*(int64_t*)&d);
case TAG_FUNCTION:
if (uintval(a) > N_BUILTINS)
return bounded_hash(((function_t*)ptr(a))->bcode, bound, ut);
return bounded_hash(((function_t*)ptr(a))->bcode, bound, oob);
return inthash(a);
case TAG_SYM:
return ((symbol_t*)ptr(a))->hash;
@ -296,39 +292,39 @@ static uptrint_t bounded_hash(value_t a, int bound, int *ut)
cv = (cvalue_t*)ptr(a);
data = cv_data(cv);
return memhash(data, cv_len(cv));
case TAG_VECTOR:
if (bound <= 0) {
h = ++(*ut) + (uptrint_t)HT_NOTFOUND;
ptrhash_put(&equal_eq_hashtable, (void*)a, (void*)h);
return h;
*oob = 1;
return 1;
}
len = vector_size(a);
for(i=0; i < len; i++) {
h = MIX(h, bounded_hash(vector_elt(a,i), bound-1, ut)+1);
h = MIX(h, bounded_hash(vector_elt(a,i), bound/2, &oob2)+1);
if (oob2)
bound/=2;
*oob = *oob || oob2;
}
return h;
case TAG_CONS:
if (bound <= 0)
if (bound <= 0) {
*oob = 1;
return 1;
return MIX(bounded_hash(car_(a), bound/2, ut),
bounded_hash(cdr_(a), bound/2, ut)+2);
// this should be able to hash long lists with greater fidelity,
// but it does not work yet.
/*
first = a;
bb = BOUNDED_HASH_BOUND;
do {
h = MIX(h, bounded_hash(car_(a), bound-1, ut));
a = cdr_(a);
bb--;
if (bb <= 0) {
*ut = 1;
ptrhash_put(&equal_eq_hashtable, (void*)first, (void*)h);
return h;
}
} while (iscons(a));
return MIX(h, bounded_hash(a, bound-1, ut));
*/
}
h = bounded_hash(car_(a), bound/2, oob);
// bounds balancing: try to share the bounds efficiently
// between the car and cdr so we can hash better when a list is
// car-shallow and cdr-deep (a common case) or vice-versa.
if (*oob)
bound/=2;
else
bound--;
h = MIX(h, bounded_hash(cdr_(a), bound, &oob2)+2);
// recursive OOB propagation. otherwise this case is slow:
// (hash '#2=('#0=(#1=(#1#) . #0#) . #2#))
*oob = *oob || oob2;
return h;
}
return 0;
}
@ -342,10 +338,8 @@ int equal_lispvalue(value_t a, value_t b)
uptrint_t hash_lispvalue(value_t a)
{
int ut=0;
uptrint_t n = bounded_hash(a, BOUNDED_HASH_BOUND, &ut);
if (ut)
htable_reset(&equal_eq_hashtable, 512);
int oob=0;
uptrint_t n = bounded_hash(a, BOUNDED_HASH_BOUND, &oob);
return n;
}

View File

@ -154,5 +154,17 @@
(hash [6 1 [2 [[3 1 [2 [1]] 3]]] 3])
(hash [6 1 [2 [[1 1 [2 [1]] 3]]] 3]))))
(assert (equal? (hash '#0=(1 . #0#))
(hash '#1=(1 1 . #1#))))
(assert (not (equal? (hash '#0=(1 1 . #0#))
(hash '#1=(1 #0# . #1#)))))
(assert (not (equal? (hash (iota 10))
(hash (iota 20)))))
(assert (not (equal? (hash (iota 41))
(hash (iota 42)))))
(princ "all tests pass\n")
#t