Don't use atomic ops in hb_cache_t
We don't care about linearizability, so unprotected int read/write are enough, no need for expensive memory barriers. It's a cache, that's all.
This commit is contained in:
parent
819faa0530
commit
e4b6d503c5
|
@ -36,7 +36,7 @@ template <unsigned int key_bits, unsigned int value_bits, unsigned int cache_bit
|
|||
struct hb_cache_t
|
||||
{
|
||||
ASSERT_STATIC (key_bits >= cache_bits);
|
||||
ASSERT_STATIC (key_bits + value_bits - cache_bits < 8 * sizeof (hb_atomic_int_t));
|
||||
ASSERT_STATIC (key_bits + value_bits - cache_bits < 8 * sizeof (unsigned int));
|
||||
|
||||
inline void clear (void)
|
||||
{
|
||||
|
@ -46,7 +46,7 @@ struct hb_cache_t
|
|||
inline bool get (unsigned int key, unsigned int *value)
|
||||
{
|
||||
unsigned int k = key & ((1<<cache_bits)-1);
|
||||
unsigned int v = hb_atomic_int_get (values[k]);
|
||||
unsigned int v = values[k];
|
||||
if ((v >> value_bits) != (key >> cache_bits))
|
||||
return false;
|
||||
}
|
||||
|
@ -57,12 +57,12 @@ struct hb_cache_t
|
|||
return false; /* Overflows */
|
||||
unsigned int k = key & ((1<<cache_bits)-1);
|
||||
unsigned int v = ((key>>cache_bits)<<value_bits) | value;
|
||||
hb_atomic_int_set (values[k], v);
|
||||
values[k] = v;
|
||||
return true;
|
||||
}
|
||||
|
||||
private:
|
||||
hb_atomic_int_t values[1<<cache_bits];
|
||||
unsigned int values[1<<cache_bits];
|
||||
};
|
||||
|
||||
typedef hb_cache_t<21, 16, 8> hb_cmap_cache_t;
|
||||
|
|
Loading…
Reference in New Issue