DPDK  24.03.0
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 #include <rte_common.h>
16 
17 #ifndef RTE_TABLE_HASH_LRU_STRATEGY
18 #define RTE_TABLE_HASH_LRU_STRATEGY 2
19 #endif
20 
21 #if RTE_TABLE_HASH_LRU_STRATEGY == 2
22 
23 #if RTE_CC_IS_GNU
24 #include <x86intrin.h>
25 #else
26 #include <emmintrin.h>
27 #include <smmintrin.h>
28 #include <xmmintrin.h>
29 #endif
30 
31 #define lru_init(bucket) \
32  { bucket->lru_list = 0x0000000100020003LLU; }
33 
34 #define lru_pos(bucket) (bucket->lru_list & 0xFFFFLLU)
35 
36 #define lru_update(bucket, mru_val) \
37 do { \
38  /* set up the masks for all possible shuffles, depends on pos */\
39  static uint64_t masks[10] = { \
40  /* Shuffle order; Make Zero (see _mm_shuffle_epi8 manual) */\
41  0x0100070605040302, 0x8080808080808080, \
42  0x0302070605040100, 0x8080808080808080, \
43  0x0504070603020100, 0x8080808080808080, \
44  0x0706050403020100, 0x8080808080808080, \
45  0x0706050403020100, 0x8080808080808080}; \
46  /* load up one register with repeats of mru-val */ \
47  uint64_t mru2 = mru_val; \
48  uint64_t mru3 = mru2 | (mru2 << 16); \
49  uint64_t lru = bucket->lru_list; \
50  /* XOR to cause the word we're looking for to go to zero */ \
51  uint64_t mru = lru ^ ((mru3 << 32) | mru3); \
52  __m128i c = _mm_cvtsi64_si128(mru); \
53  __m128i b = _mm_cvtsi64_si128(lru); \
54  /* Find the minimum value (first zero word, if it's in there) */\
55  __m128i d = _mm_minpos_epu16(c); \
56  /* Second word is the index to found word (first word is the value) */\
57  unsigned int pos = _mm_extract_epi16(d, 1); \
58  /* move the recently used location to top of list */ \
59  __m128i k = _mm_shuffle_epi8(b, *((__m128i *) &masks[2 * pos]));\
60  /* Finally, update the original list with the reordered data */ \
61  bucket->lru_list = _mm_extract_epi64(k, 0); \
62  /* Phwew! */ \
63 } while (0)
64 
65 #elif RTE_TABLE_HASH_LRU_STRATEGY == 3
66 
67 #if RTE_CC_IS_GNU
68 #include <x86intrin.h>
69 #else
70 #include <emmintrin.h>
71 #include <smmintrin.h>
72 #include <xmmintrin.h>
73 #endif
74 
75 #define lru_init(bucket) \
76  { bucket->lru_list = ~0LLU; }
77 
78 static inline int
79 f_lru_pos(uint64_t lru_list)
80 {
81  __m128i lst = _mm_set_epi64x((uint64_t)-1, lru_list);
82  __m128i min = _mm_minpos_epu16(lst);
83  return _mm_extract_epi16(min, 1);
84 }
85 #define lru_pos(bucket) f_lru_pos(bucket->lru_list)
86 
87 #define lru_update(bucket, mru_val) \
88 do { \
89  const uint64_t orvals[] = {0xFFFFLLU, 0xFFFFLLU << 16, \
90  0xFFFFLLU << 32, 0xFFFFLLU << 48, 0LLU}; \
91  const uint64_t decs[] = {0x1000100010001LLU, 0}; \
92  __m128i lru = _mm_cvtsi64_si128(bucket->lru_list); \
93  __m128i vdec = _mm_cvtsi64_si128(decs[mru_val>>2]); \
94  lru = _mm_subs_epu16(lru, vdec); \
95  bucket->lru_list = _mm_extract_epi64(lru, 0) | orvals[mru_val]; \
96 } while (0)
97 
98 #endif
99 
100 #ifdef __cplusplus
101 }
102 #endif
103 
104 #endif