Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 1 | /** |
| 2 | * \file bn_mul.h |
Paul Bakker | e0ccd0a | 2009-01-04 16:27:10 +0000 | [diff] [blame] | 3 | * |
Paul Bakker | 37ca75d | 2011-01-06 12:28:03 +0000 | [diff] [blame] | 4 | * \brief Multi-precision integer library |
| 5 | * |
Paul Bakker | 84f12b7 | 2010-07-18 10:13:04 +0000 | [diff] [blame] | 6 | * Copyright (C) 2006-2010, Brainspark B.V. |
Paul Bakker | b96f154 | 2010-07-18 20:36:00 +0000 | [diff] [blame] | 7 | * |
| 8 | * This file is part of PolarSSL (http://www.polarssl.org) |
Paul Bakker | 84f12b7 | 2010-07-18 10:13:04 +0000 | [diff] [blame] | 9 | * Lead Maintainer: Paul Bakker <polarssl_maintainer at polarssl.org> |
Paul Bakker | b96f154 | 2010-07-18 20:36:00 +0000 | [diff] [blame] | 10 | * |
Paul Bakker | 77b385e | 2009-07-28 17:23:11 +0000 | [diff] [blame] | 11 | * All rights reserved. |
Paul Bakker | e0ccd0a | 2009-01-04 16:27:10 +0000 | [diff] [blame] | 12 | * |
Paul Bakker | e0ccd0a | 2009-01-04 16:27:10 +0000 | [diff] [blame] | 13 | * This program is free software; you can redistribute it and/or modify |
| 14 | * it under the terms of the GNU General Public License as published by |
| 15 | * the Free Software Foundation; either version 2 of the License, or |
| 16 | * (at your option) any later version. |
| 17 | * |
| 18 | * This program is distributed in the hope that it will be useful, |
| 19 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 20 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 21 | * GNU General Public License for more details. |
| 22 | * |
| 23 | * You should have received a copy of the GNU General Public License along |
| 24 | * with this program; if not, write to the Free Software Foundation, Inc., |
| 25 | * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 26 | */ |
| 27 | /* |
| 28 | * Multiply source vector [s] with b, add result |
| 29 | * to destination vector [d] and set carry c. |
| 30 | * |
| 31 | * Currently supports: |
| 32 | * |
| 33 | * . IA-32 (386+) . AMD64 / EM64T |
| 34 | * . IA-32 (SSE2) . Motorola 68000 |
| 35 | * . PowerPC, 32-bit . MicroBlaze |
| 36 | * . PowerPC, 64-bit . TriCore |
| 37 | * . SPARC v8 . ARM v3+ |
| 38 | * . Alpha . MIPS32 |
| 39 | * . C, longlong . C, generic |
| 40 | */ |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 41 | #ifndef POLARSSL_BN_MUL_H |
| 42 | #define POLARSSL_BN_MUL_H |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 43 | |
Paul Bakker | 6621987 | 2012-01-22 20:38:13 +0000 | [diff] [blame] | 44 | #include "bignum.h" |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 45 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 46 | #if defined(POLARSSL_HAVE_ASM) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 47 | |
| 48 | #if defined(__GNUC__) |
| 49 | #if defined(__i386__) |
| 50 | |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 51 | #define MULADDC_INIT \ |
| 52 | asm( " \ |
| 53 | movl %%ebx, %0; \ |
| 54 | movl %5, %%esi; \ |
| 55 | movl %6, %%edi; \ |
| 56 | movl %7, %%ecx; \ |
| 57 | movl %8, %%ebx; \ |
| 58 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 59 | |
Paul Bakker | 40fe2b0 | 2009-07-08 19:49:01 +0000 | [diff] [blame] | 60 | #define MULADDC_CORE \ |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 61 | " \ |
Paul Bakker | 40fe2b0 | 2009-07-08 19:49:01 +0000 | [diff] [blame] | 62 | lodsl; \ |
| 63 | mull %%ebx; \ |
| 64 | addl %%ecx, %%eax; \ |
| 65 | adcl $0, %%edx; \ |
| 66 | addl (%%edi), %%eax; \ |
| 67 | adcl $0, %%edx; \ |
| 68 | movl %%edx, %%ecx; \ |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 69 | stosl; \ |
| 70 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 71 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 72 | #if defined(POLARSSL_HAVE_SSE2) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 73 | |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 74 | #define MULADDC_HUIT \ |
| 75 | " \ |
| 76 | movd %%ecx, %%mm1; \ |
| 77 | movd %%ebx, %%mm0; \ |
| 78 | movd (%%edi), %%mm3; \ |
| 79 | paddq %%mm3, %%mm1; \ |
| 80 | movd (%%esi), %%mm2; \ |
| 81 | pmuludq %%mm0, %%mm2; \ |
| 82 | movd 4(%%esi), %%mm4; \ |
| 83 | pmuludq %%mm0, %%mm4; \ |
| 84 | movd 8(%%esi), %%mm6; \ |
| 85 | pmuludq %%mm0, %%mm6; \ |
| 86 | movd 12(%%esi), %%mm7; \ |
| 87 | pmuludq %%mm0, %%mm7; \ |
| 88 | paddq %%mm2, %%mm1; \ |
| 89 | movd 4(%%edi), %%mm3; \ |
| 90 | paddq %%mm4, %%mm3; \ |
| 91 | movd 8(%%edi), %%mm5; \ |
| 92 | paddq %%mm6, %%mm5; \ |
| 93 | movd 12(%%edi), %%mm4; \ |
| 94 | paddq %%mm4, %%mm7; \ |
| 95 | movd %%mm1, (%%edi); \ |
| 96 | movd 16(%%esi), %%mm2; \ |
| 97 | pmuludq %%mm0, %%mm2; \ |
| 98 | psrlq $32, %%mm1; \ |
| 99 | movd 20(%%esi), %%mm4; \ |
| 100 | pmuludq %%mm0, %%mm4; \ |
| 101 | paddq %%mm3, %%mm1; \ |
| 102 | movd 24(%%esi), %%mm6; \ |
| 103 | pmuludq %%mm0, %%mm6; \ |
| 104 | movd %%mm1, 4(%%edi); \ |
| 105 | psrlq $32, %%mm1; \ |
| 106 | movd 28(%%esi), %%mm3; \ |
| 107 | pmuludq %%mm0, %%mm3; \ |
| 108 | paddq %%mm5, %%mm1; \ |
| 109 | movd 16(%%edi), %%mm5; \ |
| 110 | paddq %%mm5, %%mm2; \ |
| 111 | movd %%mm1, 8(%%edi); \ |
| 112 | psrlq $32, %%mm1; \ |
| 113 | paddq %%mm7, %%mm1; \ |
| 114 | movd 20(%%edi), %%mm5; \ |
| 115 | paddq %%mm5, %%mm4; \ |
| 116 | movd %%mm1, 12(%%edi); \ |
| 117 | psrlq $32, %%mm1; \ |
| 118 | paddq %%mm2, %%mm1; \ |
| 119 | movd 24(%%edi), %%mm5; \ |
| 120 | paddq %%mm5, %%mm6; \ |
| 121 | movd %%mm1, 16(%%edi); \ |
| 122 | psrlq $32, %%mm1; \ |
| 123 | paddq %%mm4, %%mm1; \ |
| 124 | movd 28(%%edi), %%mm5; \ |
| 125 | paddq %%mm5, %%mm3; \ |
| 126 | movd %%mm1, 20(%%edi); \ |
| 127 | psrlq $32, %%mm1; \ |
| 128 | paddq %%mm6, %%mm1; \ |
| 129 | movd %%mm1, 24(%%edi); \ |
| 130 | psrlq $32, %%mm1; \ |
| 131 | paddq %%mm3, %%mm1; \ |
| 132 | movd %%mm1, 28(%%edi); \ |
| 133 | addl $32, %%edi; \ |
| 134 | addl $32, %%esi; \ |
| 135 | psrlq $32, %%mm1; \ |
| 136 | movd %%mm1, %%ecx; \ |
| 137 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 138 | |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 139 | #define MULADDC_STOP \ |
| 140 | " \ |
| 141 | emms; \ |
| 142 | movl %4, %%ebx; \ |
| 143 | movl %%ecx, %1; \ |
| 144 | movl %%edi, %2; \ |
| 145 | movl %%esi, %3; \ |
| 146 | " \ |
| 147 | : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ |
| 148 | : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 149 | : "eax", "ecx", "edx", "esi", "edi" \ |
| 150 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 151 | |
| 152 | #else |
| 153 | |
Paul Bakker | c89cf7c | 2009-07-19 21:37:39 +0000 | [diff] [blame] | 154 | #define MULADDC_STOP \ |
| 155 | " \ |
| 156 | movl %4, %%ebx; \ |
| 157 | movl %%ecx, %1; \ |
| 158 | movl %%edi, %2; \ |
| 159 | movl %%esi, %3; \ |
| 160 | " \ |
| 161 | : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ |
| 162 | : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 163 | : "eax", "ecx", "edx", "esi", "edi" \ |
| 164 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 165 | #endif /* SSE2 */ |
| 166 | #endif /* i386 */ |
| 167 | |
| 168 | #if defined(__amd64__) || defined (__x86_64__) |
| 169 | |
Manuel Pégourié-Gonnard | def018d | 2014-01-07 17:50:46 +0100 | [diff] [blame] | 170 | #define MULADDC_INIT \ |
| 171 | asm( \ |
| 172 | " \ |
| 173 | movq %3, %%rsi; \ |
| 174 | movq %4, %%rdi; \ |
| 175 | movq %5, %%rcx; \ |
| 176 | movq %6, %%rbx; \ |
| 177 | xorq %%r8, %%r8; \ |
| 178 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 179 | |
Manuel Pégourié-Gonnard | def018d | 2014-01-07 17:50:46 +0100 | [diff] [blame] | 180 | #define MULADDC_CORE \ |
| 181 | " \ |
| 182 | movq (%%rsi), %%rax; \ |
| 183 | mulq %%rbx; \ |
| 184 | addq $8, %%rsi; \ |
| 185 | addq %%rcx, %%rax; \ |
| 186 | movq %%r8, %%rcx; \ |
| 187 | adcq $0, %%rdx; \ |
| 188 | nop; \ |
| 189 | addq %%rax, (%%rdi); \ |
| 190 | adcq %%rdx, %%rcx; \ |
| 191 | addq $8, %%rdi; \ |
| 192 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 193 | |
Manuel Pégourié-Gonnard | def018d | 2014-01-07 17:50:46 +0100 | [diff] [blame] | 194 | #define MULADDC_STOP \ |
| 195 | " \ |
| 196 | movq %%rcx, %0; \ |
| 197 | movq %%rdi, %1; \ |
| 198 | movq %%rsi, %2; \ |
| 199 | " \ |
| 200 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 201 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 202 | : "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" \ |
| 203 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 204 | |
| 205 | #endif /* AMD64 */ |
| 206 | |
| 207 | #if defined(__mc68020__) || defined(__mcpu32__) |
| 208 | |
Manuel Pégourié-Gonnard | 3b05e4c | 2014-01-10 15:30:23 +0100 | [diff] [blame] | 209 | #define MULADDC_INIT \ |
| 210 | asm( \ |
| 211 | " \ |
| 212 | movl %3, %%a2; \ |
| 213 | movl %4, %%a3; \ |
| 214 | movl %5, %%d3; \ |
| 215 | movl %6, %%d2; \ |
| 216 | moveq #0, %%d0; \ |
| 217 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 218 | |
Manuel Pégourié-Gonnard | 3b05e4c | 2014-01-10 15:30:23 +0100 | [diff] [blame] | 219 | #define MULADDC_CORE \ |
| 220 | " \ |
| 221 | movel %%a2@+, %%d1; \ |
| 222 | mulul %%d2, %%d4:%%d1; \ |
| 223 | addl %%d3, %%d1; \ |
| 224 | addxl %%d0, %%d4; \ |
| 225 | moveq #0, %%d3; \ |
| 226 | addl %%d1, %%a3@+; \ |
| 227 | addxl %%d4, %%d3; \ |
| 228 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 229 | |
Manuel Pégourié-Gonnard | 3b05e4c | 2014-01-10 15:30:23 +0100 | [diff] [blame] | 230 | #define MULADDC_STOP \ |
| 231 | " \ |
| 232 | movl %%d3, %0; \ |
| 233 | movl %%a3, %1; \ |
| 234 | movl %%a2, %2; \ |
| 235 | " \ |
| 236 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 237 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 238 | : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \ |
| 239 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 240 | |
Manuel Pégourié-Gonnard | 3b05e4c | 2014-01-10 15:30:23 +0100 | [diff] [blame] | 241 | #define MULADDC_HUIT \ |
| 242 | " \ |
| 243 | movel %%a2@+, %%d1; \ |
| 244 | mulul %%d2, %%d4:%%d1; \ |
| 245 | addxl %%d3, %%d1; \ |
| 246 | addxl %%d0, %%d4; \ |
| 247 | addl %%d1, %%a3@+; \ |
| 248 | movel %%a2@+, %%d1; \ |
| 249 | mulul %%d2, %%d3:%%d1; \ |
| 250 | addxl %%d4, %%d1; \ |
| 251 | addxl %%d0, %%d3; \ |
| 252 | addl %%d1, %%a3@+; \ |
| 253 | movel %%a2@+, %%d1; \ |
| 254 | mulul %%d2, %%d4:%%d1; \ |
| 255 | addxl %%d3, %%d1; \ |
| 256 | addxl %%d0, %%d4; \ |
| 257 | addl %%d1, %%a3@+; \ |
| 258 | movel %%a2@+, %%d1; \ |
| 259 | mulul %%d2, %%d3:%%d1; \ |
| 260 | addxl %%d4, %%d1; \ |
| 261 | addxl %%d0, %%d3; \ |
| 262 | addl %%d1, %%a3@+; \ |
| 263 | movel %%a2@+, %%d1; \ |
| 264 | mulul %%d2, %%d4:%%d1; \ |
| 265 | addxl %%d3, %%d1; \ |
| 266 | addxl %%d0, %%d4; \ |
| 267 | addl %%d1, %%a3@+; \ |
| 268 | movel %%a2@+, %%d1; \ |
| 269 | mulul %%d2, %%d3:%%d1; \ |
| 270 | addxl %%d4, %%d1; \ |
| 271 | addxl %%d0, %%d3; \ |
| 272 | addl %%d1, %%a3@+; \ |
| 273 | movel %%a2@+, %%d1; \ |
| 274 | mulul %%d2, %%d4:%%d1; \ |
| 275 | addxl %%d3, %%d1; \ |
| 276 | addxl %%d0, %%d4; \ |
| 277 | addl %%d1, %%a3@+; \ |
| 278 | movel %%a2@+, %%d1; \ |
| 279 | mulul %%d2, %%d3:%%d1; \ |
| 280 | addxl %%d4, %%d1; \ |
| 281 | addxl %%d0, %%d3; \ |
| 282 | addl %%d1, %%a3@+; \ |
| 283 | addxl %%d0, %%d3; \ |
| 284 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 285 | |
| 286 | #endif /* MC68000 */ |
| 287 | |
| 288 | #if defined(__powerpc__) || defined(__ppc__) |
| 289 | #if defined(__powerpc64__) || defined(__ppc64__) |
| 290 | |
| 291 | #if defined(__MACH__) && defined(__APPLE__) |
| 292 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 293 | #define MULADDC_INIT \ |
| 294 | asm( \ |
| 295 | " \ |
| 296 | ld r3, %3; \ |
| 297 | ld r4, %4; \ |
| 298 | ld r5, %5; \ |
| 299 | ld r6, %6; \ |
| 300 | addi r3, r3, -8; \ |
| 301 | addi r4, r4, -8; \ |
| 302 | addic r5, r5, 0; \ |
| 303 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 304 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 305 | #define MULADDC_CORE \ |
| 306 | " \ |
| 307 | ldu r7, 8(r3); \ |
| 308 | mulld r8, r7, r6; \ |
| 309 | mulhdu r9, r7, r6; \ |
| 310 | adde r8, r8, r5; \ |
| 311 | ld r7, 8(r4); \ |
| 312 | addze r5, r9; \ |
| 313 | addc r8, r8, r7; \ |
| 314 | stdu r8, 8(r4); \ |
| 315 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 316 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 317 | #define MULADDC_STOP \ |
| 318 | " \ |
| 319 | addze r5, r5; \ |
| 320 | addi r4, r4, 8; \ |
| 321 | addi r3, r3, 8; \ |
| 322 | std r5, %0; \ |
| 323 | std r4, %1; \ |
| 324 | std r3, %2; \ |
| 325 | " \ |
| 326 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 327 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 328 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| 329 | ); |
| 330 | |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 331 | |
| 332 | #else |
| 333 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 334 | #define MULADDC_INIT \ |
| 335 | asm( \ |
| 336 | " \ |
| 337 | ld %%r3, %3; \ |
| 338 | ld %%r4, %4; \ |
| 339 | ld %%r5, %5; \ |
| 340 | ld %%r6, %6; \ |
| 341 | addi %%r3, %%r3, -8; \ |
| 342 | addi %%r4, %%r4, -8; \ |
| 343 | addic %%r5, %%r5, 0; \ |
| 344 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 345 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 346 | #define MULADDC_CORE \ |
| 347 | " \ |
| 348 | ldu %%r7, 8(%%r3); \ |
| 349 | mulld %%r8, %%r7, %%r6; \ |
| 350 | mulhdu %%r9, %%r7, %%r6; \ |
| 351 | adde %%r8, %%r8, %%r5; \ |
| 352 | ld %%r7, 8(%%r4); \ |
| 353 | addze %%r5, %%r9; \ |
| 354 | addc %%r8, %%r8, %%r7; \ |
| 355 | stdu %%r8, 8(%%r4); \ |
| 356 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 357 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 358 | #define MULADDC_STOP \ |
| 359 | " \ |
| 360 | addze %%r5, %%r5; \ |
| 361 | addi %%r4, %%r4, 8; \ |
| 362 | addi %%r3, %%r3, 8; \ |
| 363 | std %%r5, %0; \ |
| 364 | std %%r4, %1; \ |
| 365 | std %%r3, %2; \ |
| 366 | " \ |
| 367 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 368 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 369 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| 370 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 371 | |
| 372 | #endif |
| 373 | |
| 374 | #else /* PPC32 */ |
| 375 | |
| 376 | #if defined(__MACH__) && defined(__APPLE__) |
| 377 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 378 | #define MULADDC_INIT \ |
| 379 | asm( \ |
| 380 | " \ |
| 381 | lwz r3, %3; \ |
| 382 | lwz r4, %4; \ |
| 383 | lwz r5, %5; \ |
| 384 | lwz r6, %6; \ |
| 385 | addi r3, r3, -4; \ |
| 386 | addi r4, r4, -4; \ |
| 387 | addic r5, r5, 0; \ |
| 388 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 389 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 390 | #define MULADDC_CORE \ |
| 391 | " \ |
| 392 | lwzu r7, 4(r3); \ |
| 393 | mullw r8, r7, r6; \ |
| 394 | mulhwu r9, r7, r6; \ |
| 395 | adde r8, r8, r5; \ |
| 396 | lwz r7, 4(r4); \ |
| 397 | addze r5, r9; \ |
| 398 | addc r8, r8, r7; \ |
| 399 | stwu r8, 4(r4); \ |
| 400 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 401 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 402 | #define MULADDC_STOP \ |
| 403 | " \ |
| 404 | addze r5, r5; \ |
| 405 | addi r4, r4, 4; \ |
| 406 | addi r3, r3, 4; \ |
| 407 | stw r5, %0; \ |
| 408 | stw r4, %1; \ |
| 409 | stw r3, %2; \ |
| 410 | " \ |
| 411 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 412 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 413 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| 414 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 415 | |
| 416 | #else |
| 417 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 418 | #define MULADDC_INIT \ |
| 419 | asm( \ |
| 420 | " \ |
| 421 | lwz %%r3, %3; \ |
| 422 | lwz %%r4, %4; \ |
| 423 | lwz %%r5, %5; \ |
| 424 | lwz %%r6, %6; \ |
| 425 | addi %%r3, %%r3, -4; \ |
| 426 | addi %%r4, %%r4, -4; \ |
| 427 | addic %%r5, %%r5, 0; \ |
| 428 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 429 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 430 | #define MULADDC_CORE \ |
| 431 | " \ |
| 432 | lwzu %%r7, 4(%%r3); \ |
| 433 | mullw %%r8, %%r7, %%r6; \ |
| 434 | mulhwu %%r9, %%r7, %%r6; \ |
| 435 | adde %%r8, %%r8, %%r5; \ |
| 436 | lwz %%r7, 4(%%r4); \ |
| 437 | addze %%r5, %%r9; \ |
| 438 | addc %%r8, %%r8, %%r7; \ |
| 439 | stwu %%r8, 4(%%r4); \ |
| 440 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 441 | |
Manuel Pégourié-Gonnard | 02d800c | 2014-01-07 19:16:48 +0100 | [diff] [blame] | 442 | #define MULADDC_STOP \ |
| 443 | " \ |
| 444 | addze %%r5, %%r5; \ |
| 445 | addi %%r4, %%r4, 4; \ |
| 446 | addi %%r3, %%r3, 4; \ |
| 447 | stw %%r5, %0; \ |
| 448 | stw %%r4, %1; \ |
| 449 | stw %%r3, %2; \ |
| 450 | " \ |
| 451 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 452 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 453 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| 454 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 455 | |
| 456 | #endif |
| 457 | |
| 458 | #endif /* PPC32 */ |
| 459 | #endif /* PPC64 */ |
| 460 | |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 461 | #if defined(__sparc__) && defined(__sparc64__) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 462 | |
| 463 | #define MULADDC_INIT \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 464 | asm( \ |
| 465 | " \ |
| 466 | ldx %3, %%o0; \ |
| 467 | ldx %4, %%o1; \ |
| 468 | ld %5, %%o2; \ |
| 469 | ld %6, %%o3; \ |
| 470 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 471 | |
| 472 | #define MULADDC_CORE \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 473 | " \ |
| 474 | ld [%%o0], %%o4; \ |
| 475 | inc 4, %%o0; \ |
| 476 | ld [%%o1], %%o5; \ |
| 477 | umul %%o3, %%o4, %%o4; \ |
| 478 | addcc %%o4, %%o2, %%o4; \ |
| 479 | rd %%y, %%g1; \ |
| 480 | addx %%g1, 0, %%g1; \ |
| 481 | addcc %%o4, %%o5, %%o4; \ |
| 482 | st %%o4, [%%o1]; \ |
| 483 | addx %%g1, 0, %%o2; \ |
| 484 | inc 4, %%o1; \ |
| 485 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 486 | |
| 487 | #define MULADDC_STOP \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 488 | " \ |
| 489 | st %%o2, %0; \ |
| 490 | stx %%o1, %1; \ |
| 491 | stx %%o0, %2; \ |
| 492 | " \ |
| 493 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 494 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 495 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| 496 | "o5" \ |
| 497 | ); |
| 498 | #endif /* SPARCv9 */ |
| 499 | |
| 500 | #if defined(__sparc__) && !defined(__sparc64__) |
| 501 | |
| 502 | #define MULADDC_INIT \ |
| 503 | asm( \ |
| 504 | " \ |
| 505 | ld %3, %%o0; \ |
| 506 | ld %4, %%o1; \ |
| 507 | ld %5, %%o2; \ |
| 508 | ld %6, %%o3; \ |
| 509 | " |
| 510 | |
| 511 | #define MULADDC_CORE \ |
| 512 | " \ |
| 513 | ld [%%o0], %%o4; \ |
| 514 | inc 4, %%o0; \ |
| 515 | ld [%%o1], %%o5; \ |
| 516 | umul %%o3, %%o4, %%o4; \ |
| 517 | addcc %%o4, %%o2, %%o4; \ |
| 518 | rd %%y, %%g1; \ |
| 519 | addx %%g1, 0, %%g1; \ |
| 520 | addcc %%o4, %%o5, %%o4; \ |
| 521 | st %%o4, [%%o1]; \ |
| 522 | addx %%g1, 0, %%o2; \ |
| 523 | inc 4, %%o1; \ |
| 524 | " |
| 525 | |
| 526 | #define MULADDC_STOP \ |
| 527 | " \ |
| 528 | st %%o2, %0; \ |
| 529 | st %%o1, %1; \ |
| 530 | st %%o0, %2; \ |
| 531 | " \ |
| 532 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 533 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 534 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| 535 | "o5" \ |
| 536 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 537 | |
| 538 | #endif /* SPARCv8 */ |
| 539 | |
| 540 | #if defined(__microblaze__) || defined(microblaze) |
| 541 | |
Manuel Pégourié-Gonnard | 1753e2f | 2014-01-10 15:35:41 +0100 | [diff] [blame] | 542 | #define MULADDC_INIT \ |
| 543 | asm( \ |
| 544 | " \ |
| 545 | lwi r3, %3; \ |
| 546 | lwi r4, %4; \ |
| 547 | lwi r5, %5; \ |
| 548 | lwi r6, %6; \ |
| 549 | andi r7, r6, 0xffff; \ |
| 550 | bsrli r6, r6, 16; \ |
| 551 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 552 | |
Manuel Pégourié-Gonnard | 1753e2f | 2014-01-10 15:35:41 +0100 | [diff] [blame] | 553 | #define MULADDC_CORE \ |
| 554 | " \ |
| 555 | lhui r8, r3, 0; \ |
| 556 | addi r3, r3, 2; \ |
| 557 | lhui r9, r3, 0; \ |
| 558 | addi r3, r3, 2; \ |
| 559 | mul r10, r9, r6; \ |
| 560 | mul r11, r8, r7; \ |
| 561 | mul r12, r9, r7; \ |
| 562 | mul r13, r8, r6; \ |
| 563 | bsrli r8, r10, 16; \ |
| 564 | bsrli r9, r11, 16; \ |
| 565 | add r13, r13, r8; \ |
| 566 | add r13, r13, r9; \ |
| 567 | bslli r10, r10, 16; \ |
| 568 | bslli r11, r11, 16; \ |
| 569 | add r12, r12, r10; \ |
| 570 | addc r13, r13, r0; \ |
| 571 | add r12, r12, r11; \ |
| 572 | addc r13, r13, r0; \ |
| 573 | lwi r10, r4, 0; \ |
| 574 | add r12, r12, r10; \ |
| 575 | addc r13, r13, r0; \ |
| 576 | add r12, r12, r5; \ |
| 577 | addc r5, r13, r0; \ |
| 578 | swi r12, r4, 0; \ |
| 579 | addi r4, r4, 4; \ |
| 580 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 581 | |
Manuel Pégourié-Gonnard | 1753e2f | 2014-01-10 15:35:41 +0100 | [diff] [blame] | 582 | #define MULADDC_STOP \ |
| 583 | " \ |
| 584 | swi r5, %0; \ |
| 585 | swi r4, %1; \ |
| 586 | swi r3, %2; \ |
| 587 | " \ |
| 588 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 589 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 590 | : "r3", "r4" "r5", "r6", "r7", "r8", \ |
| 591 | "r9", "r10", "r11", "r12", "r13" \ |
| 592 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 593 | |
| 594 | #endif /* MicroBlaze */ |
| 595 | |
| 596 | #if defined(__tricore__) |
| 597 | |
Manuel Pégourié-Gonnard | 3f687ad | 2014-01-10 15:47:50 +0100 | [diff] [blame] | 598 | #define MULADDC_INIT \ |
| 599 | asm( \ |
| 600 | " \ |
| 601 | ld.a %%a2, %3; \ |
| 602 | ld.a %%a3, %4; \ |
| 603 | ld.w %%d4, %5; \ |
| 604 | ld.w %%d1, %6; \ |
| 605 | xor %%d5, %%d5; \ |
| 606 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 607 | |
Manuel Pégourié-Gonnard | 3f687ad | 2014-01-10 15:47:50 +0100 | [diff] [blame] | 608 | #define MULADDC_CORE \ |
| 609 | " \ |
| 610 | ld.w %%d0, [%%a2+]; \ |
| 611 | madd.u %%e2, %%e4, %%d0, %%d1; \ |
| 612 | ld.w %%d0, [%%a3]; \ |
| 613 | addx %%d2, %%d2, %%d0; \ |
| 614 | addc %%d3, %%d3, 0; \ |
| 615 | mov %%d4, %%d3; \ |
| 616 | st.w [%%a3+], %%d2; \ |
| 617 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 618 | |
Manuel Pégourié-Gonnard | 3f687ad | 2014-01-10 15:47:50 +0100 | [diff] [blame] | 619 | #define MULADDC_STOP \ |
| 620 | " \ |
| 621 | st.w %0, %%d4; \ |
| 622 | st.a %1, %%a3; \ |
| 623 | st.a %2, %%a2; \ |
| 624 | " \ |
| 625 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 626 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 627 | : "d0", "d1", "e2", "d4", "a2", "a3" \ |
| 628 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 629 | |
| 630 | #endif /* TriCore */ |
| 631 | |
| 632 | #if defined(__arm__) |
| 633 | |
Paul Bakker | fc4f46f | 2013-06-24 19:23:56 +0200 | [diff] [blame] | 634 | #if defined(__thumb__) && !defined(__thumb2__) |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 635 | |
| 636 | #define MULADDC_INIT \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 637 | asm( \ |
| 638 | " \ |
| 639 | ldr r0, %3; \ |
| 640 | ldr r1, %4; \ |
| 641 | ldr r2, %5; \ |
| 642 | ldr r3, %6; \ |
| 643 | lsr r7, r3, #16; \ |
| 644 | mov r9, r7; \ |
| 645 | lsl r7, r3, #16; \ |
| 646 | lsr r7, r7, #16; \ |
| 647 | mov r8, r7; \ |
| 648 | " |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 649 | |
| 650 | #define MULADDC_CORE \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 651 | " \ |
| 652 | ldmia r0!, {r6}; \ |
| 653 | lsr r7, r6, #16; \ |
| 654 | lsl r6, r6, #16; \ |
| 655 | lsr r6, r6, #16; \ |
| 656 | mov r4, r8; \ |
| 657 | mul r4, r6; \ |
| 658 | mov r3, r9; \ |
| 659 | mul r6, r3; \ |
| 660 | mov r5, r9; \ |
| 661 | mul r5, r7; \ |
| 662 | mov r3, r8; \ |
| 663 | mul r7, r3; \ |
| 664 | lsr r3, r6, #16; \ |
| 665 | add r5, r5, r3; \ |
| 666 | lsr r3, r7, #16; \ |
| 667 | add r5, r5, r3; \ |
| 668 | add r4, r4, r2; \ |
| 669 | mov r2, #0; \ |
| 670 | adc r5, r2; \ |
| 671 | lsl r3, r6, #16; \ |
| 672 | add r4, r4, r3; \ |
| 673 | adc r5, r2; \ |
| 674 | lsl r3, r7, #16; \ |
| 675 | add r4, r4, r3; \ |
| 676 | adc r5, r2; \ |
| 677 | ldr r3, [r1]; \ |
| 678 | add r4, r4, r3; \ |
| 679 | adc r2, r5; \ |
| 680 | stmia r1!, {r4}; \ |
| 681 | " |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 682 | |
| 683 | #define MULADDC_STOP \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 684 | " \ |
| 685 | str r2, %0; \ |
| 686 | str r1, %1; \ |
| 687 | str r0, %2; \ |
| 688 | " \ |
| 689 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 690 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 691 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
Paul Bakker | eff2e6d | 2013-04-11 17:13:22 +0200 | [diff] [blame] | 692 | "r6", "r7", "r8", "r9", "cc" \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 693 | ); |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 694 | |
| 695 | #else |
Paul Bakker | a2713a3 | 2011-11-18 12:47:23 +0000 | [diff] [blame] | 696 | |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 697 | #define MULADDC_INIT \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 698 | asm( \ |
| 699 | " \ |
| 700 | ldr r0, %3; \ |
| 701 | ldr r1, %4; \ |
| 702 | ldr r2, %5; \ |
| 703 | ldr r3, %6; \ |
| 704 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 705 | |
| 706 | #define MULADDC_CORE \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 707 | " \ |
| 708 | ldr r4, [r0], #4; \ |
| 709 | mov r5, #0; \ |
| 710 | ldr r6, [r1]; \ |
| 711 | umlal r2, r5, r3, r4; \ |
| 712 | adds r7, r6, r2; \ |
| 713 | adc r2, r5, #0; \ |
| 714 | str r7, [r1], #4; \ |
| 715 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 716 | |
| 717 | #define MULADDC_STOP \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 718 | " \ |
| 719 | str r2, %0; \ |
| 720 | str r1, %1; \ |
| 721 | str r0, %2; \ |
| 722 | " \ |
| 723 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 724 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 725 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
Paul Bakker | eff2e6d | 2013-04-11 17:13:22 +0200 | [diff] [blame] | 726 | "r6", "r7", "cc" \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 727 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 728 | |
Paul Bakker | a2713a3 | 2011-11-18 12:47:23 +0000 | [diff] [blame] | 729 | #endif /* Thumb */ |
| 730 | |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 731 | #endif /* ARMv3 */ |
| 732 | |
| 733 | #if defined(__alpha__) |
| 734 | |
Manuel Pégourié-Gonnard | 5af8e64 | 2014-01-10 15:53:41 +0100 | [diff] [blame^] | 735 | #define MULADDC_INIT \ |
| 736 | asm( \ |
| 737 | " \ |
| 738 | ldq $1, %3; \ |
| 739 | ldq $2, %4; \ |
| 740 | ldq $3, %5; \ |
| 741 | ldq $4, %6; \ |
| 742 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 743 | |
Manuel Pégourié-Gonnard | 5af8e64 | 2014-01-10 15:53:41 +0100 | [diff] [blame^] | 744 | #define MULADDC_CORE \ |
| 745 | " \ |
| 746 | ldq $6, 0($1); \ |
| 747 | addq $1, 8, $1; \ |
| 748 | mulq $6, $4, $7; \ |
| 749 | umulh $6, $4, $6; \ |
| 750 | addq $7, $3, $7; \ |
| 751 | cmpult $7, $3, $3; \ |
| 752 | ldq $5, 0($2); \ |
| 753 | addq $7, $5, $7; \ |
| 754 | cmpult $7, $5, $5; \ |
| 755 | stq $7, 0($2); \ |
| 756 | addq $2, 8, $2; \ |
| 757 | addq $6, $3, $3; \ |
| 758 | addq $5, $3, $3; \ |
| 759 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 760 | |
| 761 | #define MULADDC_STOP \ |
Manuel Pégourié-Gonnard | 5af8e64 | 2014-01-10 15:53:41 +0100 | [diff] [blame^] | 762 | " \ |
| 763 | stq $3, %0; \ |
| 764 | stq $2, %1; \ |
| 765 | stq $1, %2; \ |
| 766 | " \ |
| 767 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 768 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 769 | : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \ |
| 770 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 771 | #endif /* Alpha */ |
| 772 | |
| 773 | #if defined(__mips__) |
| 774 | |
Manuel Pégourié-Gonnard | 8b1b103 | 2014-01-07 18:31:06 +0100 | [diff] [blame] | 775 | #define MULADDC_INIT \ |
| 776 | asm( \ |
| 777 | " \ |
| 778 | lw $10, %3; \ |
| 779 | lw $11, %4; \ |
| 780 | lw $12, %5; \ |
| 781 | lw $13, %6; \ |
| 782 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 783 | |
Manuel Pégourié-Gonnard | 8b1b103 | 2014-01-07 18:31:06 +0100 | [diff] [blame] | 784 | #define MULADDC_CORE \ |
| 785 | " \ |
| 786 | lw $14, 0($10); \ |
| 787 | multu $13, $14; \ |
| 788 | addi $10, $10, 4; \ |
| 789 | mflo $14; \ |
| 790 | mfhi $9; \ |
| 791 | addu $14, $12, $14; \ |
| 792 | lw $15, 0($11); \ |
| 793 | sltu $12, $14, $12; \ |
| 794 | addu $15, $14, $15; \ |
| 795 | sltu $14, $15, $14; \ |
| 796 | addu $12, $12, $9; \ |
| 797 | sw $15, 0($11); \ |
| 798 | addu $12, $12, $14; \ |
| 799 | addi $11, $11, 4; \ |
| 800 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 801 | |
Manuel Pégourié-Gonnard | 8b1b103 | 2014-01-07 18:31:06 +0100 | [diff] [blame] | 802 | #define MULADDC_STOP \ |
| 803 | " \ |
| 804 | sw $12, %0; \ |
| 805 | sw $11, %1; \ |
| 806 | sw $10, %2; \ |
| 807 | " \ |
| 808 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 809 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 810 | : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \ |
| 811 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 812 | |
| 813 | #endif /* MIPS */ |
| 814 | #endif /* GNUC */ |
| 815 | |
| 816 | #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) |
| 817 | |
| 818 | #define MULADDC_INIT \ |
| 819 | __asm mov esi, s \ |
| 820 | __asm mov edi, d \ |
| 821 | __asm mov ecx, c \ |
| 822 | __asm mov ebx, b |
| 823 | |
| 824 | #define MULADDC_CORE \ |
| 825 | __asm lodsd \ |
| 826 | __asm mul ebx \ |
| 827 | __asm add eax, ecx \ |
| 828 | __asm adc edx, 0 \ |
| 829 | __asm add eax, [edi] \ |
| 830 | __asm adc edx, 0 \ |
| 831 | __asm mov ecx, edx \ |
| 832 | __asm stosd |
| 833 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 834 | #if defined(POLARSSL_HAVE_SSE2) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 835 | |
| 836 | #define EMIT __asm _emit |
| 837 | |
| 838 | #define MULADDC_HUIT \ |
| 839 | EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ |
| 840 | EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ |
| 841 | EMIT 0x0F EMIT 0x6E EMIT 0x1F \ |
| 842 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 843 | EMIT 0x0F EMIT 0x6E EMIT 0x16 \ |
| 844 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| 845 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ |
| 846 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| 847 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ |
| 848 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| 849 | EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ |
| 850 | EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ |
| 851 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| 852 | EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ |
| 853 | EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ |
| 854 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ |
| 855 | EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ |
| 856 | EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ |
| 857 | EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ |
| 858 | EMIT 0x0F EMIT 0x7E EMIT 0x0F \ |
| 859 | EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ |
| 860 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| 861 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 862 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ |
| 863 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| 864 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 865 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ |
| 866 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| 867 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ |
| 868 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 869 | EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ |
| 870 | EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ |
| 871 | EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ |
| 872 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ |
| 873 | EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ |
| 874 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ |
| 875 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 876 | EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ |
| 877 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ |
| 878 | EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ |
| 879 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ |
| 880 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 881 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| 882 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ |
| 883 | EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ |
| 884 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ |
| 885 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 886 | EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ |
| 887 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ |
| 888 | EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ |
| 889 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ |
| 890 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 891 | EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ |
| 892 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ |
| 893 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 894 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 895 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ |
| 896 | EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ |
| 897 | EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ |
| 898 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 899 | EMIT 0x0F EMIT 0x7E EMIT 0xC9 |
| 900 | |
| 901 | #define MULADDC_STOP \ |
| 902 | EMIT 0x0F EMIT 0x77 \ |
| 903 | __asm mov c, ecx \ |
| 904 | __asm mov d, edi \ |
| 905 | __asm mov s, esi \ |
| 906 | |
| 907 | #else |
| 908 | |
| 909 | #define MULADDC_STOP \ |
| 910 | __asm mov c, ecx \ |
| 911 | __asm mov d, edi \ |
| 912 | __asm mov s, esi \ |
| 913 | |
| 914 | #endif /* SSE2 */ |
| 915 | #endif /* MSVC */ |
| 916 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 917 | #endif /* POLARSSL_HAVE_ASM */ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 918 | |
| 919 | #if !defined(MULADDC_CORE) |
Paul Bakker | 0fd018e | 2012-10-23 12:44:47 +0000 | [diff] [blame] | 920 | #if defined(POLARSSL_HAVE_UDBL) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 921 | |
| 922 | #define MULADDC_INIT \ |
| 923 | { \ |
Paul Bakker | a755ca1 | 2011-04-24 09:11:17 +0000 | [diff] [blame] | 924 | t_udbl r; \ |
| 925 | t_uint r0, r1; |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 926 | |
| 927 | #define MULADDC_CORE \ |
Paul Bakker | 5c2364c | 2012-10-01 14:41:15 +0000 | [diff] [blame] | 928 | r = *(s++) * (t_udbl) b; \ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 929 | r0 = r; \ |
| 930 | r1 = r >> biL; \ |
| 931 | r0 += c; r1 += (r0 < c); \ |
| 932 | r0 += *d; r1 += (r0 < *d); \ |
| 933 | c = r1; *(d++) = r0; |
| 934 | |
| 935 | #define MULADDC_STOP \ |
| 936 | } |
| 937 | |
| 938 | #else |
| 939 | #define MULADDC_INIT \ |
| 940 | { \ |
Paul Bakker | a755ca1 | 2011-04-24 09:11:17 +0000 | [diff] [blame] | 941 | t_uint s0, s1, b0, b1; \ |
| 942 | t_uint r0, r1, rx, ry; \ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 943 | b0 = ( b << biH ) >> biH; \ |
| 944 | b1 = ( b >> biH ); |
| 945 | |
| 946 | #define MULADDC_CORE \ |
| 947 | s0 = ( *s << biH ) >> biH; \ |
| 948 | s1 = ( *s >> biH ); s++; \ |
| 949 | rx = s0 * b1; r0 = s0 * b0; \ |
| 950 | ry = s1 * b0; r1 = s1 * b1; \ |
| 951 | r1 += ( rx >> biH ); \ |
| 952 | r1 += ( ry >> biH ); \ |
| 953 | rx <<= biH; ry <<= biH; \ |
| 954 | r0 += rx; r1 += (r0 < rx); \ |
| 955 | r0 += ry; r1 += (r0 < ry); \ |
| 956 | r0 += c; r1 += (r0 < c); \ |
| 957 | r0 += *d; r1 += (r0 < *d); \ |
| 958 | c = r1; *(d++) = r0; |
| 959 | |
| 960 | #define MULADDC_STOP \ |
| 961 | } |
| 962 | |
| 963 | #endif /* C (generic) */ |
| 964 | #endif /* C (longlong) */ |
| 965 | |
| 966 | #endif /* bn_mul.h */ |