lib/librte_lpm/rte_lpm_sse.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #ifndef _RTE_LPM_SSE_H_
   6 #define _RTE_LPM_SSE_H_
   7
   8 #include <rte_branch_prediction.h>
   9 #include <rte_byteorder.h>
  10 #include <rte_common.h>
  11 #include <rte_vect.h>
  12
  13 #ifdef __cplusplus
  14 extern "C" {
  15 #endif
  16
  17 static inline void
  18 rte_lpm_lookupx4(const struct rte_lpm *lpm, xmm_t ip, uint32_t hop[4],
  19         uint32_t defv)
  20 {
  21         __m128i i24;
  22         rte_xmm_t i8;
  23         uint32_t tbl[4];
  24         uint64_t idx, pt, pt2;
  25         const uint32_t *ptbl;
  26
  27         const __m128i mask8 =
  28                 _mm_set_epi32(UINT8_MAX, UINT8_MAX, UINT8_MAX, UINT8_MAX);
  29
  30         /*
  31          * RTE_LPM_VALID_EXT_ENTRY_BITMASK for 2 LPM entries
  32          * as one 64-bit value (0x0300000003000000).
  33          */
  34         const uint64_t mask_xv =
  35                 ((uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK |
  36                 (uint64_t)RTE_LPM_VALID_EXT_ENTRY_BITMASK << 32);
  37
  38         /*
  39          * RTE_LPM_LOOKUP_SUCCESS for 2 LPM entries
  40          * as one 64-bit value (0x0100000001000000).
  41          */
  42         const uint64_t mask_v =
  43                 ((uint64_t)RTE_LPM_LOOKUP_SUCCESS |
  44                 (uint64_t)RTE_LPM_LOOKUP_SUCCESS << 32);
  45
  46         /* get 4 indexes for tbl24[]. */
  47         i24 = _mm_srli_epi32(ip, CHAR_BIT);
  48
  49         /* extract values from tbl24[] */
  50         idx = _mm_cvtsi128_si64(i24);
  51         /* With -O0 option, gcc 4.8 - 5.4 fails to fold sizeof() into a constant */
  52         i24 = _mm_srli_si128(i24, /* sizeof(uint64_t) */ 8);
  53
  54         ptbl = (const uint32_t *)&lpm->tbl24[(uint32_t)idx];
  55         tbl[0] = *ptbl;
  56         ptbl = (const uint32_t *)&lpm->tbl24[idx >> 32];
  57         tbl[1] = *ptbl;
  58
  59         idx = _mm_cvtsi128_si64(i24);
  60
  61         ptbl = (const uint32_t *)&lpm->tbl24[(uint32_t)idx];
  62         tbl[2] = *ptbl;
  63         ptbl = (const uint32_t *)&lpm->tbl24[idx >> 32];
  64         tbl[3] = *ptbl;
  65
  66         /* get 4 indexes for tbl8[]. */
  67         i8.x = _mm_and_si128(ip, mask8);
  68
  69         pt = (uint64_t)tbl[0] |
  70                 (uint64_t)tbl[1] << 32;
  71         pt2 = (uint64_t)tbl[2] |
  72                 (uint64_t)tbl[3] << 32;
  73
  74         /* search successfully finished for all 4 IP addresses. */
  75         if (likely((pt & mask_xv) == mask_v) &&
  76                         likely((pt2 & mask_xv) == mask_v)) {
  77                 *(uint64_t *)hop = pt & RTE_LPM_MASKX4_RES;
  78                 *(uint64_t *)(hop + 2) = pt2 & RTE_LPM_MASKX4_RES;
  79                 return;
  80         }
  81
  82         if (unlikely((pt & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
  83                         RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
  84                 i8.u32[0] = i8.u32[0] +
  85                         (uint8_t)tbl[0] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
  86                 ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[0]];
  87                 tbl[0] = *ptbl;
  88         }
  89         if (unlikely((pt >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
  90                         RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
  91                 i8.u32[1] = i8.u32[1] +
  92                         (uint8_t)tbl[1] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
  93                 ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[1]];
  94                 tbl[1] = *ptbl;
  95         }
  96         if (unlikely((pt2 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
  97                         RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
  98                 i8.u32[2] = i8.u32[2] +
  99                         (uint8_t)tbl[2] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 100                 ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[2]];
 101                 tbl[2] = *ptbl;
 102         }
 103         if (unlikely((pt2 >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
 104                         RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
 105                 i8.u32[3] = i8.u32[3] +
 106                         (uint8_t)tbl[3] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
 107                 ptbl = (const uint32_t *)&lpm->tbl8[i8.u32[3]];
 108                 tbl[3] = *ptbl;
 109         }
 110
 111         hop[0] = (tbl[0] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[0] & 0x00FFFFFF : defv;
 112         hop[1] = (tbl[1] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[1] & 0x00FFFFFF : defv;
 113         hop[2] = (tbl[2] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[2] & 0x00FFFFFF : defv;
 114         hop[3] = (tbl[3] & RTE_LPM_LOOKUP_SUCCESS) ? tbl[3] & 0x00FFFFFF : defv;
 115 }
 116
 117 #ifdef __cplusplus
 118 }
 119 #endif
 120
 121 #endif /* _RTE_LPM_SSE_H_ */