lib/librte_table/rte_lru_x86.h

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #ifndef __INCLUDE_RTE_LRU_X86_H__
  35 #define __INCLUDE_RTE_LRU_X86_H__
  36
  37 #ifdef __cplusplus
  38 extern "C" {
  39 #endif
  40
  41 #include <stdint.h>
  42
  43 #ifdef __INTEL_COMPILER
  44 #define GCC_VERSION (0)
  45 #else
  46 #define GCC_VERSION (__GNUC__ * 10000+__GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
  47 #endif
  48
  49 #ifndef RTE_TABLE_HASH_LRU_STRATEGY
  50 #ifdef __SSE4_2__
  51 #define RTE_TABLE_HASH_LRU_STRATEGY                        2
  52 #else /* if no SSE, use simple scalar version */
  53 #define RTE_TABLE_HASH_LRU_STRATEGY                        1
  54 #endif
  55 #endif
  56
  57 #if RTE_TABLE_HASH_LRU_STRATEGY == 2
  58
  59 #if GCC_VERSION > 40306
  60 #include <x86intrin.h>
  61 #else
  62 #include <emmintrin.h>
  63 #include <smmintrin.h>
  64 #include <xmmintrin.h>
  65 #endif
  66
  67 #define lru_init(bucket)                                                \
  68         { bucket->lru_list = 0x0000000100020003LLU; }
  69
  70 #define lru_pos(bucket) (bucket->lru_list & 0xFFFFLLU)
  71
  72 #define lru_update(bucket, mru_val)                                     \
  73 do {                                                                    \
  74         /* set up the masks for all possible shuffles, depends on pos */\
  75         static uint64_t masks[10] = {                                   \
  76                 /* Shuffle order; Make Zero (see _mm_shuffle_epi8 manual) */\
  77                 0x0100070605040302, 0x8080808080808080,                 \
  78                 0x0302070605040100, 0x8080808080808080,                 \
  79                 0x0504070603020100, 0x8080808080808080,                 \
  80                 0x0706050403020100, 0x8080808080808080,                 \
  81                 0x0706050403020100, 0x8080808080808080};                \
  82         /* load up one register with repeats of mru-val  */             \
  83         uint64_t mru2 = mru_val;                                        \
  84         uint64_t mru3 = mru2 | (mru2 << 16);                            \
  85         uint64_t lru = bucket->lru_list;                                \
  86         /* XOR to cause the word we're looking for to go to zero */     \
  87         uint64_t mru = lru ^ ((mru3 << 32) | mru3);                     \
  88         __m128i c = _mm_cvtsi64_si128(mru);                             \
  89         __m128i b = _mm_cvtsi64_si128(lru);                             \
  90         /* Find the minimum value (first zero word, if it's in there) */\
  91         __m128i d = _mm_minpos_epu16(c);                                \
  92         /* Second word is the index to found word (first word is the value) */\
  93         unsigned int pos = _mm_extract_epi16(d, 1);                     \
  94         /* move the recently used location to top of list */            \
  95         __m128i k = _mm_shuffle_epi8(b, *((__m128i *) &masks[2 * pos]));\
  96         /* Finally, update the original list with the reordered data */ \
  97         bucket->lru_list = _mm_extract_epi64(k, 0);                     \
  98         /* Phwew! */                                                    \
  99 } while (0)
 100
 101 #elif RTE_TABLE_HASH_LRU_STRATEGY == 3
 102
 103 #if GCC_VERSION > 40306
 104 #include <x86intrin.h>
 105 #else
 106 #include <emmintrin.h>
 107 #include <smmintrin.h>
 108 #include <xmmintrin.h>
 109 #endif
 110
 111 #define lru_init(bucket)                                                \
 112         { bucket->lru_list = ~0LLU; }
 113
 114 static inline int
 115 f_lru_pos(uint64_t lru_list)
 116 {
 117         __m128i lst = _mm_set_epi64x((uint64_t)-1, lru_list);
 118         __m128i min = _mm_minpos_epu16(lst);
 119         return _mm_extract_epi16(min, 1);
 120 }
 121 #define lru_pos(bucket) f_lru_pos(bucket->lru_list)
 122
 123 #define lru_update(bucket, mru_val)                                     \
 124 do {                                                                    \
 125         const uint64_t orvals[] = {0xFFFFLLU, 0xFFFFLLU << 16,          \
 126                 0xFFFFLLU << 32, 0xFFFFLLU << 48, 0LLU};                \
 127         const uint64_t decs[] = {0x1000100010001LLU, 0};                \
 128         __m128i lru = _mm_cvtsi64_si128(bucket->lru_list);              \
 129         __m128i vdec = _mm_cvtsi64_si128(decs[mru_val>>2]);             \
 130         lru = _mm_subs_epu16(lru, vdec);                                \
 131         bucket->lru_list = _mm_extract_epi64(lru, 0) | orvals[mru_val]; \
 132 } while (0)
 133
 134 #endif
 135
 136 #ifdef __cplusplus
 137 }
 138 #endif
 139
 140 #endif