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