#define L1_CACHE_BYTES 32 /* Two 16-byte sub-blocks per line. */
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))
#define L1_CACHE_BYTES 32 /* Two 16-byte sub-blocks per line. */
#define L1_CACHE_ALIGN(x) (((x)+(L1_CACHE_BYTES-1))&~(L1_CACHE_BYTES-1))