1 // Copyright 2016 The Go Authors. All rights reserved. 2 // Use of this source code is governed by a BSD-style 3 // license that can be found in the LICENSE file. 4 5 #include "textflag.h" 6 7 // SHA256 block routine. See sha256block.go for Go equivalent. 8 // 9 // The algorithm is detailed in FIPS 180-4: 10 // 11 // http://csrc.nist.gov/publications/fips/fips180-4/fips-180-4.pdf 12 // 13 // Wt = Mt; for 0 <= t <= 15 14 // Wt = SIGMA1(Wt-2) + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 63 15 // 16 // a = H0 17 // b = H1 18 // c = H2 19 // d = H3 20 // e = H4 21 // f = H5 22 // g = H6 23 // h = H7 24 // 25 // for t = 0 to 63 { 26 // T1 = h + BIGSIGMA1(e) + Ch(e,f,g) + Kt + Wt 27 // T2 = BIGSIGMA0(a) + Maj(a,b,c) 28 // h = g 29 // g = f 30 // f = e 31 // e = d + T1 32 // d = c 33 // c = b 34 // b = a 35 // a = T1 + T2 36 // } 37 // 38 // H0 = a + H0 39 // H1 = b + H1 40 // H2 = c + H2 41 // H3 = d + H3 42 // H4 = e + H4 43 // H5 = f + H5 44 // H6 = g + H6 45 // H7 = h + H7 46 47 // Wt = Mt; for 0 <= t <= 15 48 #define MSGSCHEDULE0(index) \ 49 MOVWZ (index*4)(R26), R7; \ 50 RLWNM $24, R7, $-1, R11; \ 51 RLWMI $8, R7, $0x00FF0000, R11; \ 52 RLWMI $8, R7, $0x000000FF, R11; \ 53 MOVWZ R11, R7; \ 54 MOVWZ R7, (index*4)(R27) 55 56 // Wt = SIGMA1(Wt-2) + Wt-7 + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 63 57 // SIGMA0(x) = ROTR(7,x) XOR ROTR(18,x) XOR SHR(3,x) 58 // SIGMA1(x) = ROTR(17,x) XOR ROTR(19,x) XOR SHR(10,x) 59 #define MSGSCHEDULE1(index) \ 60 MOVWZ ((index-2)*4)(R27), R7; \ 61 MOVWZ R7, R9; \ 62 RLWNM $32-17, R7, $-1, R7; \ 63 MOVWZ R9, R10; \ 64 RLWNM $32-19, R9, $-1, R9; \ 65 SRW $10, R10; \ 66 MOVWZ ((index-15)*4)(R27), R8; \ 67 XOR R9, R7; \ 68 MOVWZ R8, R9; \ 69 XOR R10, R7; \ 70 RLWNM $32-7, R8, $-1, R8; \ 71 MOVWZ R9, R10; \ 72 SRW $3, R10; \ 73 RLWNM $32-18, R9, $-1, R9; \ 74 MOVWZ ((index-7)*4)(R27), R11; \ 75 ADD R11, R7; \ 76 XOR R9, R8; \ 77 XOR R10, R8; \ 78 MOVWZ ((index-16)*4)(R27), R11; \ 79 ADD R11, R8; \ 80 ADD R8, R7; \ 81 MOVWZ R7, ((index)*4)(R27) 82 83 // T1 = h + BIGSIGMA1(e) + Ch(e, f, g) + Kt + Wt 84 // BIGSIGMA1(x) = ROTR(6,x) XOR ROTR(11,x) XOR ROTR(25,x) 85 // Ch(x, y, z) = (x AND y) XOR (NOT x AND z) 86 #define SHA256T1(const, e, f, g, h) \ 87 ADD R7, h; \ 88 MOVWZ e, R7; \ 89 ADD $const, h; \ 90 MOVWZ e, R9; \ 91 RLWNM $32-6, R7, $-1, R7; \ 92 MOVWZ e, R10; \ 93 RLWNM $32-11, R9, $-1, R9; \ 94 XOR R9, R7; \ 95 MOVWZ e, R9; \ 96 RLWNM $32-25, R10, $-1, R10; \ 97 AND f, R9; \ 98 XOR R7, R10; \ 99 MOVWZ e, R7; \ 100 NOR R7, R7, R7; \ 101 ADD R10, h; \ 102 AND g, R7; \ 103 XOR R9, R7; \ 104 ADD h, R7 105 106 // T2 = BIGSIGMA0(a) + Maj(a, b, c) 107 // BIGSIGMA0(x) = ROTR(2,x) XOR ROTR(13,x) XOR ROTR(22,x) 108 // Maj(x, y, z) = (x AND y) XOR (x AND z) XOR (y AND z) 109 #define SHA256T2(a, b, c) \ 110 MOVWZ a, R28; \ 111 MOVWZ c, R8; \ 112 RLWNM $32-2, R28, $-1, R28; \ 113 MOVWZ a, R10; \ 114 AND b, R8; \ 115 RLWNM $32-13, R10, $-1, R10; \ 116 MOVWZ a, R9; \ 117 AND c, R9; \ 118 XOR R10, R28; \ 119 XOR R9, R8; \ 120 MOVWZ a, R10; \ 121 MOVWZ b, R9; \ 122 RLWNM $32-22, R10, $-1, R10; \ 123 AND a, R9; \ 124 XOR R9, R8; \ 125 XOR R10, R28; \ 126 ADD R28, R8 127 128 // Calculate T1 and T2, then e = d + T1 and a = T1 + T2. 129 // The values for e and a are stored in d and h, ready for rotation. 130 #define SHA256ROUND(index, const, a, b, c, d, e, f, g, h) \ 131 SHA256T1(const, e, f, g, h); \ 132 SHA256T2(a, b, c); \ 133 MOVWZ R8, h; \ 134 ADD R7, d; \ 135 ADD R7, h 136 137 #define SHA256ROUND0(index, const, a, b, c, d, e, f, g, h) \ 138 MSGSCHEDULE0(index); \ 139 SHA256ROUND(index, const, a, b, c, d, e, f, g, h) 140 141 #define SHA256ROUND1(index, const, a, b, c, d, e, f, g, h) \ 142 MSGSCHEDULE1(index); \ 143 SHA256ROUND(index, const, a, b, c, d, e, f, g, h) 144 145 // func block(dig *digest, p []byte) 146 TEXT block(SB),0,$296-32 147 MOVD p_base+8(FP), R26 148 MOVD p_len+16(FP), R29 149 SRD $6, R29 150 SLD $6, R29 151 152 ADD R26, R29, R28 153 154 MOVD R28, 256(R1) 155 CMP R26, R28 156 BEQ end 157 158 MOVD dig+0(FP), R27 159 MOVWZ (0*4)(R27), R14 // a = H0 160 MOVWZ (1*4)(R27), R15 // b = H1 161 MOVWZ (2*4)(R27), R16 // c = H2 162 MOVWZ (3*4)(R27), R17 // d = H3 163 MOVWZ (4*4)(R27), R18 // e = H4 164 MOVWZ (5*4)(R27), R19 // f = H5 165 MOVWZ (6*4)(R27), R20 // g = H6 166 MOVWZ (7*4)(R27), R21 // h = H7 167 168 loop: 169 MOVD R1, R27 // R27: message schedule 170 171 SHA256ROUND0(0, 0x428a2f98, R14, R15, R16, R17, R18, R19, R20, R21) 172 SHA256ROUND0(1, 0x71374491, R21, R14, R15, R16, R17, R18, R19, R20) 173 SHA256ROUND0(2, 0xb5c0fbcf, R20, R21, R14, R15, R16, R17, R18, R19) 174 SHA256ROUND0(3, 0xe9b5dba5, R19, R20, R21, R14, R15, R16, R17, R18) 175 SHA256ROUND0(4, 0x3956c25b, R18, R19, R20, R21, R14, R15, R16, R17) 176 SHA256ROUND0(5, 0x59f111f1, R17, R18, R19, R20, R21, R14, R15, R16) 177 SHA256ROUND0(6, 0x923f82a4, R16, R17, R18, R19, R20, R21, R14, R15) 178 SHA256ROUND0(7, 0xab1c5ed5, R15, R16, R17, R18, R19, R20, R21, R14) 179 SHA256ROUND0(8, 0xd807aa98, R14, R15, R16, R17, R18, R19, R20, R21) 180 SHA256ROUND0(9, 0x12835b01, R21, R14, R15, R16, R17, R18, R19, R20) 181 SHA256ROUND0(10, 0x243185be, R20, R21, R14, R15, R16, R17, R18, R19) 182 SHA256ROUND0(11, 0x550c7dc3, R19, R20, R21, R14, R15, R16, R17, R18) 183 SHA256ROUND0(12, 0x72be5d74, R18, R19, R20, R21, R14, R15, R16, R17) 184 SHA256ROUND0(13, 0x80deb1fe, R17, R18, R19, R20, R21, R14, R15, R16) 185 SHA256ROUND0(14, 0x9bdc06a7, R16, R17, R18, R19, R20, R21, R14, R15) 186 SHA256ROUND0(15, 0xc19bf174, R15, R16, R17, R18, R19, R20, R21, R14) 187 188 SHA256ROUND1(16, 0xe49b69c1, R14, R15, R16, R17, R18, R19, R20, R21) 189 SHA256ROUND1(17, 0xefbe4786, R21, R14, R15, R16, R17, R18, R19, R20) 190 SHA256ROUND1(18, 0x0fc19dc6, R20, R21, R14, R15, R16, R17, R18, R19) 191 SHA256ROUND1(19, 0x240ca1cc, R19, R20, R21, R14, R15, R16, R17, R18) 192 SHA256ROUND1(20, 0x2de92c6f, R18, R19, R20, R21, R14, R15, R16, R17) 193 SHA256ROUND1(21, 0x4a7484aa, R17, R18, R19, R20, R21, R14, R15, R16) 194 SHA256ROUND1(22, 0x5cb0a9dc, R16, R17, R18, R19, R20, R21, R14, R15) 195 SHA256ROUND1(23, 0x76f988da, R15, R16, R17, R18, R19, R20, R21, R14) 196 SHA256ROUND1(24, 0x983e5152, R14, R15, R16, R17, R18, R19, R20, R21) 197 SHA256ROUND1(25, 0xa831c66d, R21, R14, R15, R16, R17, R18, R19, R20) 198 SHA256ROUND1(26, 0xb00327c8, R20, R21, R14, R15, R16, R17, R18, R19) 199 SHA256ROUND1(27, 0xbf597fc7, R19, R20, R21, R14, R15, R16, R17, R18) 200 SHA256ROUND1(28, 0xc6e00bf3, R18, R19, R20, R21, R14, R15, R16, R17) 201 SHA256ROUND1(29, 0xd5a79147, R17, R18, R19, R20, R21, R14, R15, R16) 202 SHA256ROUND1(30, 0x06ca6351, R16, R17, R18, R19, R20, R21, R14, R15) 203 SHA256ROUND1(31, 0x14292967, R15, R16, R17, R18, R19, R20, R21, R14) 204 SHA256ROUND1(32, 0x27b70a85, R14, R15, R16, R17, R18, R19, R20, R21) 205 SHA256ROUND1(33, 0x2e1b2138, R21, R14, R15, R16, R17, R18, R19, R20) 206 SHA256ROUND1(34, 0x4d2c6dfc, R20, R21, R14, R15, R16, R17, R18, R19) 207 SHA256ROUND1(35, 0x53380d13, R19, R20, R21, R14, R15, R16, R17, R18) 208 SHA256ROUND1(36, 0x650a7354, R18, R19, R20, R21, R14, R15, R16, R17) 209 SHA256ROUND1(37, 0x766a0abb, R17, R18, R19, R20, R21, R14, R15, R16) 210 SHA256ROUND1(38, 0x81c2c92e, R16, R17, R18, R19, R20, R21, R14, R15) 211 SHA256ROUND1(39, 0x92722c85, R15, R16, R17, R18, R19, R20, R21, R14) 212 SHA256ROUND1(40, 0xa2bfe8a1, R14, R15, R16, R17, R18, R19, R20, R21) 213 SHA256ROUND1(41, 0xa81a664b, R21, R14, R15, R16, R17, R18, R19, R20) 214 SHA256ROUND1(42, 0xc24b8b70, R20, R21, R14, R15, R16, R17, R18, R19) 215 SHA256ROUND1(43, 0xc76c51a3, R19, R20, R21, R14, R15, R16, R17, R18) 216 SHA256ROUND1(44, 0xd192e819, R18, R19, R20, R21, R14, R15, R16, R17) 217 SHA256ROUND1(45, 0xd6990624, R17, R18, R19, R20, R21, R14, R15, R16) 218 SHA256ROUND1(46, 0xf40e3585, R16, R17, R18, R19, R20, R21, R14, R15) 219 SHA256ROUND1(47, 0x106aa070, R15, R16, R17, R18, R19, R20, R21, R14) 220 SHA256ROUND1(48, 0x19a4c116, R14, R15, R16, R17, R18, R19, R20, R21) 221 SHA256ROUND1(49, 0x1e376c08, R21, R14, R15, R16, R17, R18, R19, R20) 222 SHA256ROUND1(50, 0x2748774c, R20, R21, R14, R15, R16, R17, R18, R19) 223 SHA256ROUND1(51, 0x34b0bcb5, R19, R20, R21, R14, R15, R16, R17, R18) 224 SHA256ROUND1(52, 0x391c0cb3, R18, R19, R20, R21, R14, R15, R16, R17) 225 SHA256ROUND1(53, 0x4ed8aa4a, R17, R18, R19, R20, R21, R14, R15, R16) 226 SHA256ROUND1(54, 0x5b9cca4f, R16, R17, R18, R19, R20, R21, R14, R15) 227 SHA256ROUND1(55, 0x682e6ff3, R15, R16, R17, R18, R19, R20, R21, R14) 228 SHA256ROUND1(56, 0x748f82ee, R14, R15, R16, R17, R18, R19, R20, R21) 229 SHA256ROUND1(57, 0x78a5636f, R21, R14, R15, R16, R17, R18, R19, R20) 230 SHA256ROUND1(58, 0x84c87814, R20, R21, R14, R15, R16, R17, R18, R19) 231 SHA256ROUND1(59, 0x8cc70208, R19, R20, R21, R14, R15, R16, R17, R18) 232 SHA256ROUND1(60, 0x90befffa, R18, R19, R20, R21, R14, R15, R16, R17) 233 SHA256ROUND1(61, 0xa4506ceb, R17, R18, R19, R20, R21, R14, R15, R16) 234 SHA256ROUND1(62, 0xbef9a3f7, R16, R17, R18, R19, R20, R21, R14, R15) 235 SHA256ROUND1(63, 0xc67178f2, R15, R16, R17, R18, R19, R20, R21, R14) 236 237 MOVD dig+0(FP), R27 238 MOVWZ (0*4)(R27), R11 239 ADD R11, R14 // H0 = a + H0 240 MOVWZ R14, (0*4)(R27) 241 MOVWZ (1*4)(R27), R11 242 ADD R11, R15 // H1 = b + H1 243 MOVWZ R15, (1*4)(R27) 244 MOVWZ (2*4)(R27), R11 245 ADD R11, R16 // H2 = c + H2 246 MOVWZ R16, (2*4)(R27) 247 MOVWZ (3*4)(R27), R11 248 ADD R11, R17 // H3 = d + H3 249 MOVWZ R17, (3*4)(R27) 250 MOVWZ (4*4)(R27), R11 251 ADD R11, R18 // H4 = e + H4 252 MOVWZ R18, (4*4)(R27) 253 MOVWZ (5*4)(R27), R11 254 ADD R11, R19 // H5 = f + H5 255 MOVWZ R19, (5*4)(R27) 256 MOVWZ (6*4)(R27), R11 257 ADD R11, R20 // H6 = g + H6 258 MOVWZ R20, (6*4)(R27) 259 MOVWZ (7*4)(R27), R11 260 ADD R11, R21 // H7 = h + H7 261 MOVWZ R21, (7*4)(R27) 262 263 ADD $64, R26 264 MOVD 256(R1), R11 265 CMPU R26, R11 266 BLT loop 267 268 end: 269 RET 270
