| 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
| 2 | /* |
| 3 | * SM4 helper macros for Crypto Extensions |
| 4 | * Copyright (C) 2022 Tianjia Zhang <tianjia.zhang@linux.alibaba.com> |
| 5 | */ |
| 6 | |
| 7 | #define SM4_PREPARE(ptr) \ |
| 8 | ld1 {v24.16b-v27.16b}, [ptr], #64; \ |
| 9 | ld1 {v28.16b-v31.16b}, [ptr]; |
| 10 | |
| 11 | #define SM4_CRYPT_BLK_BE(b0) \ |
| 12 | sm4e b0.4s, v24.4s; \ |
| 13 | sm4e b0.4s, v25.4s; \ |
| 14 | sm4e b0.4s, v26.4s; \ |
| 15 | sm4e b0.4s, v27.4s; \ |
| 16 | sm4e b0.4s, v28.4s; \ |
| 17 | sm4e b0.4s, v29.4s; \ |
| 18 | sm4e b0.4s, v30.4s; \ |
| 19 | sm4e b0.4s, v31.4s; \ |
| 20 | rev64 b0.4s, b0.4s; \ |
| 21 | ext b0.16b, b0.16b, b0.16b, #8; \ |
| 22 | rev32 b0.16b, b0.16b; |
| 23 | |
| 24 | #define SM4_CRYPT_BLK(b0) \ |
| 25 | rev32 b0.16b, b0.16b; \ |
| 26 | SM4_CRYPT_BLK_BE(b0); |
| 27 | |
| 28 | #define SM4_CRYPT_BLK2_BE(b0, b1) \ |
| 29 | sm4e b0.4s, v24.4s; \ |
| 30 | sm4e b1.4s, v24.4s; \ |
| 31 | sm4e b0.4s, v25.4s; \ |
| 32 | sm4e b1.4s, v25.4s; \ |
| 33 | sm4e b0.4s, v26.4s; \ |
| 34 | sm4e b1.4s, v26.4s; \ |
| 35 | sm4e b0.4s, v27.4s; \ |
| 36 | sm4e b1.4s, v27.4s; \ |
| 37 | sm4e b0.4s, v28.4s; \ |
| 38 | sm4e b1.4s, v28.4s; \ |
| 39 | sm4e b0.4s, v29.4s; \ |
| 40 | sm4e b1.4s, v29.4s; \ |
| 41 | sm4e b0.4s, v30.4s; \ |
| 42 | sm4e b1.4s, v30.4s; \ |
| 43 | sm4e b0.4s, v31.4s; \ |
| 44 | sm4e b1.4s, v31.4s; \ |
| 45 | rev64 b0.4s, b0.4s; \ |
| 46 | rev64 b1.4s, b1.4s; \ |
| 47 | ext b0.16b, b0.16b, b0.16b, #8; \ |
| 48 | ext b1.16b, b1.16b, b1.16b, #8; \ |
| 49 | rev32 b0.16b, b0.16b; \ |
| 50 | rev32 b1.16b, b1.16b; \ |
| 51 | |
| 52 | #define SM4_CRYPT_BLK2(b0, b1) \ |
| 53 | rev32 b0.16b, b0.16b; \ |
| 54 | rev32 b1.16b, b1.16b; \ |
| 55 | SM4_CRYPT_BLK2_BE(b0, b1); |
| 56 | |
| 57 | #define SM4_CRYPT_BLK4_BE(b0, b1, b2, b3) \ |
| 58 | sm4e b0.4s, v24.4s; \ |
| 59 | sm4e b1.4s, v24.4s; \ |
| 60 | sm4e b2.4s, v24.4s; \ |
| 61 | sm4e b3.4s, v24.4s; \ |
| 62 | sm4e b0.4s, v25.4s; \ |
| 63 | sm4e b1.4s, v25.4s; \ |
| 64 | sm4e b2.4s, v25.4s; \ |
| 65 | sm4e b3.4s, v25.4s; \ |
| 66 | sm4e b0.4s, v26.4s; \ |
| 67 | sm4e b1.4s, v26.4s; \ |
| 68 | sm4e b2.4s, v26.4s; \ |
| 69 | sm4e b3.4s, v26.4s; \ |
| 70 | sm4e b0.4s, v27.4s; \ |
| 71 | sm4e b1.4s, v27.4s; \ |
| 72 | sm4e b2.4s, v27.4s; \ |
| 73 | sm4e b3.4s, v27.4s; \ |
| 74 | sm4e b0.4s, v28.4s; \ |
| 75 | sm4e b1.4s, v28.4s; \ |
| 76 | sm4e b2.4s, v28.4s; \ |
| 77 | sm4e b3.4s, v28.4s; \ |
| 78 | sm4e b0.4s, v29.4s; \ |
| 79 | sm4e b1.4s, v29.4s; \ |
| 80 | sm4e b2.4s, v29.4s; \ |
| 81 | sm4e b3.4s, v29.4s; \ |
| 82 | sm4e b0.4s, v30.4s; \ |
| 83 | sm4e b1.4s, v30.4s; \ |
| 84 | sm4e b2.4s, v30.4s; \ |
| 85 | sm4e b3.4s, v30.4s; \ |
| 86 | sm4e b0.4s, v31.4s; \ |
| 87 | sm4e b1.4s, v31.4s; \ |
| 88 | sm4e b2.4s, v31.4s; \ |
| 89 | sm4e b3.4s, v31.4s; \ |
| 90 | rev64 b0.4s, b0.4s; \ |
| 91 | rev64 b1.4s, b1.4s; \ |
| 92 | rev64 b2.4s, b2.4s; \ |
| 93 | rev64 b3.4s, b3.4s; \ |
| 94 | ext b0.16b, b0.16b, b0.16b, #8; \ |
| 95 | ext b1.16b, b1.16b, b1.16b, #8; \ |
| 96 | ext b2.16b, b2.16b, b2.16b, #8; \ |
| 97 | ext b3.16b, b3.16b, b3.16b, #8; \ |
| 98 | rev32 b0.16b, b0.16b; \ |
| 99 | rev32 b1.16b, b1.16b; \ |
| 100 | rev32 b2.16b, b2.16b; \ |
| 101 | rev32 b3.16b, b3.16b; |
| 102 | |
| 103 | #define SM4_CRYPT_BLK4(b0, b1, b2, b3) \ |
| 104 | rev32 b0.16b, b0.16b; \ |
| 105 | rev32 b1.16b, b1.16b; \ |
| 106 | rev32 b2.16b, b2.16b; \ |
| 107 | rev32 b3.16b, b3.16b; \ |
| 108 | SM4_CRYPT_BLK4_BE(b0, b1, b2, b3); |
| 109 | |
| 110 | #define SM4_CRYPT_BLK8_BE(b0, b1, b2, b3, b4, b5, b6, b7) \ |
| 111 | sm4e b0.4s, v24.4s; \ |
| 112 | sm4e b1.4s, v24.4s; \ |
| 113 | sm4e b2.4s, v24.4s; \ |
| 114 | sm4e b3.4s, v24.4s; \ |
| 115 | sm4e b4.4s, v24.4s; \ |
| 116 | sm4e b5.4s, v24.4s; \ |
| 117 | sm4e b6.4s, v24.4s; \ |
| 118 | sm4e b7.4s, v24.4s; \ |
| 119 | sm4e b0.4s, v25.4s; \ |
| 120 | sm4e b1.4s, v25.4s; \ |
| 121 | sm4e b2.4s, v25.4s; \ |
| 122 | sm4e b3.4s, v25.4s; \ |
| 123 | sm4e b4.4s, v25.4s; \ |
| 124 | sm4e b5.4s, v25.4s; \ |
| 125 | sm4e b6.4s, v25.4s; \ |
| 126 | sm4e b7.4s, v25.4s; \ |
| 127 | sm4e b0.4s, v26.4s; \ |
| 128 | sm4e b1.4s, v26.4s; \ |
| 129 | sm4e b2.4s, v26.4s; \ |
| 130 | sm4e b3.4s, v26.4s; \ |
| 131 | sm4e b4.4s, v26.4s; \ |
| 132 | sm4e b5.4s, v26.4s; \ |
| 133 | sm4e b6.4s, v26.4s; \ |
| 134 | sm4e b7.4s, v26.4s; \ |
| 135 | sm4e b0.4s, v27.4s; \ |
| 136 | sm4e b1.4s, v27.4s; \ |
| 137 | sm4e b2.4s, v27.4s; \ |
| 138 | sm4e b3.4s, v27.4s; \ |
| 139 | sm4e b4.4s, v27.4s; \ |
| 140 | sm4e b5.4s, v27.4s; \ |
| 141 | sm4e b6.4s, v27.4s; \ |
| 142 | sm4e b7.4s, v27.4s; \ |
| 143 | sm4e b0.4s, v28.4s; \ |
| 144 | sm4e b1.4s, v28.4s; \ |
| 145 | sm4e b2.4s, v28.4s; \ |
| 146 | sm4e b3.4s, v28.4s; \ |
| 147 | sm4e b4.4s, v28.4s; \ |
| 148 | sm4e b5.4s, v28.4s; \ |
| 149 | sm4e b6.4s, v28.4s; \ |
| 150 | sm4e b7.4s, v28.4s; \ |
| 151 | sm4e b0.4s, v29.4s; \ |
| 152 | sm4e b1.4s, v29.4s; \ |
| 153 | sm4e b2.4s, v29.4s; \ |
| 154 | sm4e b3.4s, v29.4s; \ |
| 155 | sm4e b4.4s, v29.4s; \ |
| 156 | sm4e b5.4s, v29.4s; \ |
| 157 | sm4e b6.4s, v29.4s; \ |
| 158 | sm4e b7.4s, v29.4s; \ |
| 159 | sm4e b0.4s, v30.4s; \ |
| 160 | sm4e b1.4s, v30.4s; \ |
| 161 | sm4e b2.4s, v30.4s; \ |
| 162 | sm4e b3.4s, v30.4s; \ |
| 163 | sm4e b4.4s, v30.4s; \ |
| 164 | sm4e b5.4s, v30.4s; \ |
| 165 | sm4e b6.4s, v30.4s; \ |
| 166 | sm4e b7.4s, v30.4s; \ |
| 167 | sm4e b0.4s, v31.4s; \ |
| 168 | sm4e b1.4s, v31.4s; \ |
| 169 | sm4e b2.4s, v31.4s; \ |
| 170 | sm4e b3.4s, v31.4s; \ |
| 171 | sm4e b4.4s, v31.4s; \ |
| 172 | sm4e b5.4s, v31.4s; \ |
| 173 | sm4e b6.4s, v31.4s; \ |
| 174 | sm4e b7.4s, v31.4s; \ |
| 175 | rev64 b0.4s, b0.4s; \ |
| 176 | rev64 b1.4s, b1.4s; \ |
| 177 | rev64 b2.4s, b2.4s; \ |
| 178 | rev64 b3.4s, b3.4s; \ |
| 179 | rev64 b4.4s, b4.4s; \ |
| 180 | rev64 b5.4s, b5.4s; \ |
| 181 | rev64 b6.4s, b6.4s; \ |
| 182 | rev64 b7.4s, b7.4s; \ |
| 183 | ext b0.16b, b0.16b, b0.16b, #8; \ |
| 184 | ext b1.16b, b1.16b, b1.16b, #8; \ |
| 185 | ext b2.16b, b2.16b, b2.16b, #8; \ |
| 186 | ext b3.16b, b3.16b, b3.16b, #8; \ |
| 187 | ext b4.16b, b4.16b, b4.16b, #8; \ |
| 188 | ext b5.16b, b5.16b, b5.16b, #8; \ |
| 189 | ext b6.16b, b6.16b, b6.16b, #8; \ |
| 190 | ext b7.16b, b7.16b, b7.16b, #8; \ |
| 191 | rev32 b0.16b, b0.16b; \ |
| 192 | rev32 b1.16b, b1.16b; \ |
| 193 | rev32 b2.16b, b2.16b; \ |
| 194 | rev32 b3.16b, b3.16b; \ |
| 195 | rev32 b4.16b, b4.16b; \ |
| 196 | rev32 b5.16b, b5.16b; \ |
| 197 | rev32 b6.16b, b6.16b; \ |
| 198 | rev32 b7.16b, b7.16b; |
| 199 | |
| 200 | #define SM4_CRYPT_BLK8(b0, b1, b2, b3, b4, b5, b6, b7) \ |
| 201 | rev32 b0.16b, b0.16b; \ |
| 202 | rev32 b1.16b, b1.16b; \ |
| 203 | rev32 b2.16b, b2.16b; \ |
| 204 | rev32 b3.16b, b3.16b; \ |
| 205 | rev32 b4.16b, b4.16b; \ |
| 206 | rev32 b5.16b, b5.16b; \ |
| 207 | rev32 b6.16b, b6.16b; \ |
| 208 | rev32 b7.16b, b7.16b; \ |
| 209 | SM4_CRYPT_BLK8_BE(b0, b1, b2, b3, b4, b5, b6, b7); |
| 210 | |