Remove cache, rotate arena_get start scan index
This yields a ~23% overall performance improvment. It operates on the idea that always iterating from the beginning will constantly get slower as more objects are allocated- the free objects have higher and higher indices. Instead, rotate around the arena so we
This commit is contained in:
58
src/arena.c
58
src/arena.c
@@ -20,7 +20,7 @@ struct ucl_arena *ucl_arena_create(size_t object_size, size_t capacity) {
|
|||||||
arena->capacity = capacity;
|
arena->capacity = capacity;
|
||||||
arena->objects = malloc(capacity * object_size);
|
arena->objects = malloc(capacity * object_size);
|
||||||
arena->used_map = malloc(used_map_size);
|
arena->used_map = malloc(used_map_size);
|
||||||
arena->next_free = 0;
|
arena->last_allocated_index = 0;
|
||||||
arena->stats.used = 0;
|
arena->stats.used = 0;
|
||||||
arena->stats.freed = 0;
|
arena->stats.freed = 0;
|
||||||
|
|
||||||
@@ -50,45 +50,25 @@ void ucl_arena_map(struct ucl_arena *arena, void (*map_function)(struct ucl_aren
|
|||||||
|
|
||||||
|
|
||||||
void *ucl_arena_get(struct ucl_arena *arena) {
|
void *ucl_arena_get(struct ucl_arena *arena) {
|
||||||
int used_map_ints = DIV_ROUND_UP(arena->capacity, INT_BITS);
|
size_t index = arena->last_allocated_index;
|
||||||
|
do {
|
||||||
|
// TODO: Restructure for speed with ffs
|
||||||
if (arena->next_free != -1) {
|
index++;
|
||||||
// Cache hit
|
if (index == arena->capacity) {
|
||||||
/* unsigned int int_index = arena->next_free / INT_BITS; */
|
index = 0;
|
||||||
/* unsigned int bit_index = arena->next_free % INT_BITS; */
|
|
||||||
|
|
||||||
unsigned int int_index = arena->next_free >> 5;
|
|
||||||
unsigned int bit_index = arena->next_free & 31;
|
|
||||||
|
|
||||||
arena->used_map[int_index] |= 1 << bit_index;
|
|
||||||
|
|
||||||
void *ret = (char *) arena->objects + (arena->next_free * arena->object_size);
|
|
||||||
arena->next_free = -1;
|
|
||||||
|
|
||||||
arena->stats.used += 1;
|
|
||||||
return ret;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Cache miss
|
|
||||||
for (int i = 0; i < used_map_ints; i++ ) {
|
|
||||||
// TODO: Maybe keep a cache of available used_map_ints
|
|
||||||
int map = arena->used_map[i];
|
|
||||||
if (!(~map)) {
|
|
||||||
continue;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
unsigned int bit_index = __builtin_ffs(~map) - 1;
|
size_t int_index = index / INT_BITS;
|
||||||
unsigned int index = bit_index + INT_BITS * i;
|
size_t bit_index = index % INT_BITS;
|
||||||
if (index >= arena->capacity) {
|
|
||||||
// This might happen in the last used_map_int when (capacity % int_bits != 0)
|
|
||||||
return NULL;
|
|
||||||
}
|
|
||||||
arena->used_map[i] |= 1 << bit_index;
|
|
||||||
arena->stats.used += 1;
|
|
||||||
|
|
||||||
return (char *) arena->objects + (index * arena->object_size);
|
if (!(arena->used_map[int_index] & 1 << bit_index)) {
|
||||||
}
|
arena->used_map[int_index] |= 1 << bit_index;
|
||||||
|
arena->last_allocated_index = index;
|
||||||
|
arena->stats.used += 1;
|
||||||
|
return (char *) arena->objects + (index * arena->object_size);
|
||||||
|
}
|
||||||
|
|
||||||
|
} while (index != arena->last_allocated_index);
|
||||||
|
|
||||||
return NULL;
|
return NULL;
|
||||||
}
|
}
|
||||||
@@ -113,10 +93,6 @@ void ucl_arena_put(struct ucl_arena *arena, void *object) {
|
|||||||
|
|
||||||
arena->used_map[int_index] &= ~(1 << bit_index);
|
arena->used_map[int_index] &= ~(1 << bit_index);
|
||||||
|
|
||||||
/* if (arena->next_free == -1) { */
|
|
||||||
arena->next_free = index;
|
|
||||||
/* } */
|
|
||||||
|
|
||||||
arena->stats.used -= 1;
|
arena->stats.used -= 1;
|
||||||
arena->stats.freed += 1;
|
arena->stats.freed += 1;
|
||||||
}
|
}
|
||||||
|
|||||||
@@ -56,7 +56,7 @@ struct ucl_arena {
|
|||||||
void *objects;
|
void *objects;
|
||||||
int *used_map;
|
int *used_map;
|
||||||
/** Cache of size 1 **/
|
/** Cache of size 1 **/
|
||||||
int next_free;
|
size_t last_allocated_index;
|
||||||
|
|
||||||
struct {
|
struct {
|
||||||
size_t used;
|
size_t used;
|
||||||
|
|||||||
Reference in New Issue
Block a user