rte_lru_arm64.h

/* SPDX-License-Identifier: BSD-3-Clause
 * Copyright(c) 2017 Cavium, Inc
 */

#ifndef __RTE_LRU_ARM64_H__
#define __RTE_LRU_ARM64_H__

#include <stdint.h>
#include <rte_vect.h>
#include <rte_bitops.h>

#ifdef __cplusplus
extern "C" {
#endif

#ifndef RTE_TABLE_HASH_LRU_STRATEGY
#ifdef __ARM_NEON
#define RTE_TABLE_HASH_LRU_STRATEGY                        3
#else /* if no NEON, use simple scalar version */
#define RTE_TABLE_HASH_LRU_STRATEGY                        1
#endif
#endif

#if RTE_TABLE_HASH_LRU_STRATEGY == 3

#define lru_init(bucket)                        \
    { bucket->lru_list = ~0LLU; }

static inline int
f_lru_pos(uint64_t lru_list)
{
    /* Compare the vector to zero vector */
    uint16x4_t lru_vec = vld1_u16((uint16_t *)&lru_list);
    uint16x4_t min_vec = vmov_n_u16(vminv_u16(lru_vec));
    uint64_t mask = vget_lane_u64(vreinterpret_u64_u16(
            vceq_u16(min_vec, lru_vec)), 0);
    return rte_clz64(mask) >> 4;
}
#define lru_pos(bucket) f_lru_pos(bucket->lru_list)

#define lru_update(bucket, mru_val)                 \
do {                                    \
    const uint64_t _orvals[] = {0xFFFFLLU, 0xFFFFLLU << 16,     \
        0xFFFFLLU << 32, 0xFFFFLLU << 48, 0LLU};        \
    const uint64_t _decs[] = {0x1000100010001LLU, 0};       \
    uint64x1_t _lru = vdup_n_u64(bucket->lru_list);         \
    uint64x1_t _vdec = vdup_n_u64(_decs[mru_val>>2]);           \
    bucket->lru_list = vget_lane_u64(vreinterpret_u64_u16(      \
                vsub_u16(vreinterpret_u16_u64(_lru),    \
                    vreinterpret_u16_u64(_vdec))),  \
                0);                 \
    bucket->lru_list |= _orvals[mru_val];               \
} while (0)

#endif

#ifdef __cplusplus
}
#endif

#endif