lib/librte_table/rte_lru_x86.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #ifndef __INCLUDE_RTE_LRU_X86_H__
   6 #define __INCLUDE_RTE_LRU_X86_H__
   7
   8 #ifdef __cplusplus
   9 extern "C" {
  10 #endif
  11
  12 #include <stdint.h>
  13
  14 #include <rte_config.h>
  15
  16 #ifndef RTE_TABLE_HASH_LRU_STRATEGY
  17 #define RTE_TABLE_HASH_LRU_STRATEGY                        2
  18 #endif
  19
  20 #if RTE_TABLE_HASH_LRU_STRATEGY == 2
  21
  22 #if RTE_CC_IS_GNU && (GCC_VERSION > 40306)
  23 #include <x86intrin.h>
  24 #else
  25 #include <emmintrin.h>
  26 #include <smmintrin.h>
  27 #include <xmmintrin.h>
  28 #endif
  29
  30 #define lru_init(bucket)                                                \
  31         { bucket->lru_list = 0x0000000100020003LLU; }
  32
  33 #define lru_pos(bucket) (bucket->lru_list & 0xFFFFLLU)
  34
  35 #define lru_update(bucket, mru_val)                                     \
  36 do {                                                                    \
  37         /* set up the masks for all possible shuffles, depends on pos */\
  38         static uint64_t masks[10] = {                                   \
  39                 /* Shuffle order; Make Zero (see _mm_shuffle_epi8 manual) */\
  40                 0x0100070605040302, 0x8080808080808080,                 \
  41                 0x0302070605040100, 0x8080808080808080,                 \
  42                 0x0504070603020100, 0x8080808080808080,                 \
  43                 0x0706050403020100, 0x8080808080808080,                 \
  44                 0x0706050403020100, 0x8080808080808080};                \
  45         /* load up one register with repeats of mru-val  */             \
  46         uint64_t mru2 = mru_val;                                        \
  47         uint64_t mru3 = mru2 | (mru2 << 16);                            \
  48         uint64_t lru = bucket->lru_list;                                \
  49         /* XOR to cause the word we're looking for to go to zero */     \
  50         uint64_t mru = lru ^ ((mru3 << 32) | mru3);                     \
  51         __m128i c = _mm_cvtsi64_si128(mru);                             \
  52         __m128i b = _mm_cvtsi64_si128(lru);                             \
  53         /* Find the minimum value (first zero word, if it's in there) */\
  54         __m128i d = _mm_minpos_epu16(c);                                \
  55         /* Second word is the index to found word (first word is the value) */\
  56         unsigned int pos = _mm_extract_epi16(d, 1);                     \
  57         /* move the recently used location to top of list */            \
  58         __m128i k = _mm_shuffle_epi8(b, *((__m128i *) &masks[2 * pos]));\
  59         /* Finally, update the original list with the reordered data */ \
  60         bucket->lru_list = _mm_extract_epi64(k, 0);                     \
  61         /* Phwew! */                                                    \
  62 } while (0)
  63
  64 #elif RTE_TABLE_HASH_LRU_STRATEGY == 3
  65
  66 #if RTE_CC_IS_GNU && (GCC_VERSION > 40306)
  67 #include <x86intrin.h>
  68 #else
  69 #include <emmintrin.h>
  70 #include <smmintrin.h>
  71 #include <xmmintrin.h>
  72 #endif
  73
  74 #define lru_init(bucket)                                                \
  75         { bucket->lru_list = ~0LLU; }
  76
  77 static inline int
  78 f_lru_pos(uint64_t lru_list)
  79 {
  80         __m128i lst = _mm_set_epi64x((uint64_t)-1, lru_list);
  81         __m128i min = _mm_minpos_epu16(lst);
  82         return _mm_extract_epi16(min, 1);
  83 }
  84 #define lru_pos(bucket) f_lru_pos(bucket->lru_list)
  85
  86 #define lru_update(bucket, mru_val)                                     \
  87 do {                                                                    \
  88         const uint64_t orvals[] = {0xFFFFLLU, 0xFFFFLLU << 16,          \
  89                 0xFFFFLLU << 32, 0xFFFFLLU << 48, 0LLU};                \
  90         const uint64_t decs[] = {0x1000100010001LLU, 0};                \
  91         __m128i lru = _mm_cvtsi64_si128(bucket->lru_list);              \
  92         __m128i vdec = _mm_cvtsi64_si128(decs[mru_val>>2]);             \
  93         lru = _mm_subs_epu16(lru, vdec);                                \
  94         bucket->lru_list = _mm_extract_epi64(lru, 0) | orvals[mru_val]; \
  95 } while (0)
  96
  97 #endif
  98
  99 #ifdef __cplusplus
 100 }
 101 #endif
 102
 103 #endif