DPDK  24.11.2
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 #include <stdint.h>
9 
10 #include <rte_config.h>
11 #include <rte_common.h>
12 
13 #ifndef RTE_TABLE_HASH_LRU_STRATEGY
14 #define RTE_TABLE_HASH_LRU_STRATEGY 2
15 #endif
16 
17 #if RTE_TABLE_HASH_LRU_STRATEGY == 2
18 
19 #if RTE_CC_IS_GNU
20 #include <x86intrin.h>
21 #else
22 #include <emmintrin.h>
23 #include <smmintrin.h>
24 #include <xmmintrin.h>
25 #endif
26 
27 #define lru_init(bucket) \
28  { bucket->lru_list = 0x0000000100020003LLU; }
29 
30 #define lru_pos(bucket) (bucket->lru_list & 0xFFFFLLU)
31 
32 #define lru_update(bucket, mru_val) \
33 do { \
34  /* set up the masks for all possible shuffles, depends on pos */\
35  static uint64_t masks[10] = { \
36  /* Shuffle order; Make Zero (see _mm_shuffle_epi8 manual) */\
37  0x0100070605040302, 0x8080808080808080, \
38  0x0302070605040100, 0x8080808080808080, \
39  0x0504070603020100, 0x8080808080808080, \
40  0x0706050403020100, 0x8080808080808080, \
41  0x0706050403020100, 0x8080808080808080}; \
42  /* load up one register with repeats of mru-val */ \
43  uint64_t mru2 = mru_val; \
44  uint64_t mru3 = mru2 | (mru2 << 16); \
45  uint64_t lru = bucket->lru_list; \
46  /* XOR to cause the word we're looking for to go to zero */ \
47  uint64_t mru = lru ^ ((mru3 << 32) | mru3); \
48  __m128i c = _mm_cvtsi64_si128(mru); \
49  __m128i b = _mm_cvtsi64_si128(lru); \
50  /* Find the minimum value (first zero word, if it's in there) */\
51  __m128i d = _mm_minpos_epu16(c); \
52  /* Second word is the index to found word (first word is the value) */\
53  unsigned int pos = _mm_extract_epi16(d, 1); \
54  /* move the recently used location to top of list */ \
55  __m128i k = _mm_shuffle_epi8(b, *((__m128i *) &masks[2 * pos]));\
56  /* Finally, update the original list with the reordered data */ \
57  bucket->lru_list = _mm_extract_epi64(k, 0); \
58  /* Phwew! */ \
59 } while (0)
60 
61 #elif RTE_TABLE_HASH_LRU_STRATEGY == 3
62 
63 #if RTE_CC_IS_GNU
64 #include <x86intrin.h>
65 #else
66 #include <emmintrin.h>
67 #include <smmintrin.h>
68 #include <xmmintrin.h>
69 #endif
70 
71 #define lru_init(bucket) \
72  { bucket->lru_list = ~0LLU; }
73 
74 static inline int
75 f_lru_pos(uint64_t lru_list)
76 {
77  __m128i lst = _mm_set_epi64x((uint64_t)-1, lru_list);
78  __m128i min = _mm_minpos_epu16(lst);
79  return _mm_extract_epi16(min, 1);
80 }
81 #define lru_pos(bucket) f_lru_pos(bucket->lru_list)
82 
83 #define lru_update(bucket, mru_val) \
84 do { \
85  const uint64_t orvals[] = {0xFFFFLLU, 0xFFFFLLU << 16, \
86  0xFFFFLLU << 32, 0xFFFFLLU << 48, 0LLU}; \
87  const uint64_t decs[] = {0x1000100010001LLU, 0}; \
88  __m128i lru = _mm_cvtsi64_si128(bucket->lru_list); \
89  __m128i vdec = _mm_cvtsi64_si128(decs[mru_val>>2]); \
90  lru = _mm_subs_epu16(lru, vdec); \
91  bucket->lru_list = _mm_extract_epi64(lru, 0) | orvals[mru_val]; \
92 } while (0)
93 
94 #endif
95 
96 #endif