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 | |
| 209 | #define MULADDC_INIT \ |
| 210 | asm( "movl %0, %%a2 " :: "m" (s)); \ |
| 211 | asm( "movl %0, %%a3 " :: "m" (d)); \ |
| 212 | asm( "movl %0, %%d3 " :: "m" (c)); \ |
| 213 | asm( "movl %0, %%d2 " :: "m" (b)); \ |
| 214 | asm( "moveq #0, %d0 " ); |
| 215 | |
| 216 | #define MULADDC_CORE \ |
| 217 | asm( "movel %a2@+, %d1 " ); \ |
| 218 | asm( "mulul %d2, %d4:%d1 " ); \ |
| 219 | asm( "addl %d3, %d1 " ); \ |
| 220 | asm( "addxl %d0, %d4 " ); \ |
| 221 | asm( "moveq #0, %d3 " ); \ |
| 222 | asm( "addl %d1, %a3@+ " ); \ |
| 223 | asm( "addxl %d4, %d3 " ); |
| 224 | |
| 225 | #define MULADDC_STOP \ |
| 226 | asm( "movl %%d3, %0 " : "=m" (c)); \ |
| 227 | asm( "movl %%a3, %0 " : "=m" (d)); \ |
| 228 | asm( "movl %%a2, %0 " : "=m" (s) :: \ |
| 229 | "d0", "d1", "d2", "d3", "d4", "a2", "a3" ); |
| 230 | |
| 231 | #define MULADDC_HUIT \ |
| 232 | asm( "movel %a2@+, %d1 " ); \ |
| 233 | asm( "mulul %d2, %d4:%d1 " ); \ |
| 234 | asm( "addxl %d3, %d1 " ); \ |
| 235 | asm( "addxl %d0, %d4 " ); \ |
| 236 | asm( "addl %d1, %a3@+ " ); \ |
| 237 | asm( "movel %a2@+, %d1 " ); \ |
| 238 | asm( "mulul %d2, %d3:%d1 " ); \ |
| 239 | asm( "addxl %d4, %d1 " ); \ |
| 240 | asm( "addxl %d0, %d3 " ); \ |
| 241 | asm( "addl %d1, %a3@+ " ); \ |
| 242 | asm( "movel %a2@+, %d1 " ); \ |
| 243 | asm( "mulul %d2, %d4:%d1 " ); \ |
| 244 | asm( "addxl %d3, %d1 " ); \ |
| 245 | asm( "addxl %d0, %d4 " ); \ |
| 246 | asm( "addl %d1, %a3@+ " ); \ |
| 247 | asm( "movel %a2@+, %d1 " ); \ |
| 248 | asm( "mulul %d2, %d3:%d1 " ); \ |
| 249 | asm( "addxl %d4, %d1 " ); \ |
| 250 | asm( "addxl %d0, %d3 " ); \ |
| 251 | asm( "addl %d1, %a3@+ " ); \ |
| 252 | asm( "movel %a2@+, %d1 " ); \ |
| 253 | asm( "mulul %d2, %d4:%d1 " ); \ |
| 254 | asm( "addxl %d3, %d1 " ); \ |
| 255 | asm( "addxl %d0, %d4 " ); \ |
| 256 | asm( "addl %d1, %a3@+ " ); \ |
| 257 | asm( "movel %a2@+, %d1 " ); \ |
| 258 | asm( "mulul %d2, %d3:%d1 " ); \ |
| 259 | asm( "addxl %d4, %d1 " ); \ |
| 260 | asm( "addxl %d0, %d3 " ); \ |
| 261 | asm( "addl %d1, %a3@+ " ); \ |
| 262 | asm( "movel %a2@+, %d1 " ); \ |
| 263 | asm( "mulul %d2, %d4:%d1 " ); \ |
| 264 | asm( "addxl %d3, %d1 " ); \ |
| 265 | asm( "addxl %d0, %d4 " ); \ |
| 266 | asm( "addl %d1, %a3@+ " ); \ |
| 267 | asm( "movel %a2@+, %d1 " ); \ |
| 268 | asm( "mulul %d2, %d3:%d1 " ); \ |
| 269 | asm( "addxl %d4, %d1 " ); \ |
| 270 | asm( "addxl %d0, %d3 " ); \ |
| 271 | asm( "addl %d1, %a3@+ " ); \ |
| 272 | asm( "addxl %d0, %d3 " ); |
| 273 | |
| 274 | #endif /* MC68000 */ |
| 275 | |
| 276 | #if defined(__powerpc__) || defined(__ppc__) |
| 277 | #if defined(__powerpc64__) || defined(__ppc64__) |
| 278 | |
| 279 | #if defined(__MACH__) && defined(__APPLE__) |
| 280 | |
| 281 | #define MULADDC_INIT \ |
| 282 | asm( "ld r3, %0 " :: "m" (s)); \ |
| 283 | asm( "ld r4, %0 " :: "m" (d)); \ |
| 284 | asm( "ld r5, %0 " :: "m" (c)); \ |
| 285 | asm( "ld r6, %0 " :: "m" (b)); \ |
| 286 | asm( "addi r3, r3, -8 " ); \ |
| 287 | asm( "addi r4, r4, -8 " ); \ |
| 288 | asm( "addic r5, r5, 0 " ); |
| 289 | |
| 290 | #define MULADDC_CORE \ |
| 291 | asm( "ldu r7, 8(r3) " ); \ |
| 292 | asm( "mulld r8, r7, r6 " ); \ |
| 293 | asm( "mulhdu r9, r7, r6 " ); \ |
| 294 | asm( "adde r8, r8, r5 " ); \ |
| 295 | asm( "ld r7, 8(r4) " ); \ |
| 296 | asm( "addze r5, r9 " ); \ |
| 297 | asm( "addc r8, r8, r7 " ); \ |
| 298 | asm( "stdu r8, 8(r4) " ); |
| 299 | |
| 300 | #define MULADDC_STOP \ |
| 301 | asm( "addze r5, r5 " ); \ |
| 302 | asm( "addi r4, r4, 8 " ); \ |
| 303 | asm( "addi r3, r3, 8 " ); \ |
| 304 | asm( "std r5, %0 " : "=m" (c)); \ |
| 305 | asm( "std r4, %0 " : "=m" (d)); \ |
| 306 | asm( "std r3, %0 " : "=m" (s) :: \ |
| 307 | "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); |
| 308 | |
| 309 | #else |
| 310 | |
| 311 | #define MULADDC_INIT \ |
| 312 | asm( "ld %%r3, %0 " :: "m" (s)); \ |
| 313 | asm( "ld %%r4, %0 " :: "m" (d)); \ |
| 314 | asm( "ld %%r5, %0 " :: "m" (c)); \ |
| 315 | asm( "ld %%r6, %0 " :: "m" (b)); \ |
| 316 | asm( "addi %r3, %r3, -8 " ); \ |
| 317 | asm( "addi %r4, %r4, -8 " ); \ |
| 318 | asm( "addic %r5, %r5, 0 " ); |
| 319 | |
| 320 | #define MULADDC_CORE \ |
| 321 | asm( "ldu %r7, 8(%r3) " ); \ |
| 322 | asm( "mulld %r8, %r7, %r6 " ); \ |
| 323 | asm( "mulhdu %r9, %r7, %r6 " ); \ |
| 324 | asm( "adde %r8, %r8, %r5 " ); \ |
| 325 | asm( "ld %r7, 8(%r4) " ); \ |
| 326 | asm( "addze %r5, %r9 " ); \ |
| 327 | asm( "addc %r8, %r8, %r7 " ); \ |
| 328 | asm( "stdu %r8, 8(%r4) " ); |
| 329 | |
| 330 | #define MULADDC_STOP \ |
| 331 | asm( "addze %r5, %r5 " ); \ |
| 332 | asm( "addi %r4, %r4, 8 " ); \ |
| 333 | asm( "addi %r3, %r3, 8 " ); \ |
| 334 | asm( "std %%r5, %0 " : "=m" (c)); \ |
| 335 | asm( "std %%r4, %0 " : "=m" (d)); \ |
| 336 | asm( "std %%r3, %0 " : "=m" (s) :: \ |
| 337 | "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); |
| 338 | |
| 339 | #endif |
| 340 | |
| 341 | #else /* PPC32 */ |
| 342 | |
| 343 | #if defined(__MACH__) && defined(__APPLE__) |
| 344 | |
| 345 | #define MULADDC_INIT \ |
| 346 | asm( "lwz r3, %0 " :: "m" (s)); \ |
| 347 | asm( "lwz r4, %0 " :: "m" (d)); \ |
| 348 | asm( "lwz r5, %0 " :: "m" (c)); \ |
| 349 | asm( "lwz r6, %0 " :: "m" (b)); \ |
| 350 | asm( "addi r3, r3, -4 " ); \ |
| 351 | asm( "addi r4, r4, -4 " ); \ |
| 352 | asm( "addic r5, r5, 0 " ); |
| 353 | |
| 354 | #define MULADDC_CORE \ |
| 355 | asm( "lwzu r7, 4(r3) " ); \ |
| 356 | asm( "mullw r8, r7, r6 " ); \ |
| 357 | asm( "mulhwu r9, r7, r6 " ); \ |
| 358 | asm( "adde r8, r8, r5 " ); \ |
| 359 | asm( "lwz r7, 4(r4) " ); \ |
| 360 | asm( "addze r5, r9 " ); \ |
| 361 | asm( "addc r8, r8, r7 " ); \ |
| 362 | asm( "stwu r8, 4(r4) " ); |
| 363 | |
| 364 | #define MULADDC_STOP \ |
| 365 | asm( "addze r5, r5 " ); \ |
| 366 | asm( "addi r4, r4, 4 " ); \ |
| 367 | asm( "addi r3, r3, 4 " ); \ |
| 368 | asm( "stw r5, %0 " : "=m" (c)); \ |
| 369 | asm( "stw r4, %0 " : "=m" (d)); \ |
| 370 | asm( "stw r3, %0 " : "=m" (s) :: \ |
| 371 | "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); |
| 372 | |
| 373 | #else |
| 374 | |
| 375 | #define MULADDC_INIT \ |
| 376 | asm( "lwz %%r3, %0 " :: "m" (s)); \ |
| 377 | asm( "lwz %%r4, %0 " :: "m" (d)); \ |
| 378 | asm( "lwz %%r5, %0 " :: "m" (c)); \ |
| 379 | asm( "lwz %%r6, %0 " :: "m" (b)); \ |
| 380 | asm( "addi %r3, %r3, -4 " ); \ |
| 381 | asm( "addi %r4, %r4, -4 " ); \ |
| 382 | asm( "addic %r5, %r5, 0 " ); |
| 383 | |
| 384 | #define MULADDC_CORE \ |
| 385 | asm( "lwzu %r7, 4(%r3) " ); \ |
| 386 | asm( "mullw %r8, %r7, %r6 " ); \ |
| 387 | asm( "mulhwu %r9, %r7, %r6 " ); \ |
| 388 | asm( "adde %r8, %r8, %r5 " ); \ |
| 389 | asm( "lwz %r7, 4(%r4) " ); \ |
| 390 | asm( "addze %r5, %r9 " ); \ |
| 391 | asm( "addc %r8, %r8, %r7 " ); \ |
| 392 | asm( "stwu %r8, 4(%r4) " ); |
| 393 | |
| 394 | #define MULADDC_STOP \ |
| 395 | asm( "addze %r5, %r5 " ); \ |
| 396 | asm( "addi %r4, %r4, 4 " ); \ |
| 397 | asm( "addi %r3, %r3, 4 " ); \ |
| 398 | asm( "stw %%r5, %0 " : "=m" (c)); \ |
| 399 | asm( "stw %%r4, %0 " : "=m" (d)); \ |
| 400 | asm( "stw %%r3, %0 " : "=m" (s) :: \ |
| 401 | "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); |
| 402 | |
| 403 | #endif |
| 404 | |
| 405 | #endif /* PPC32 */ |
| 406 | #endif /* PPC64 */ |
| 407 | |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 408 | #if defined(__sparc__) && defined(__sparc64__) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 409 | |
| 410 | #define MULADDC_INIT \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 411 | asm( \ |
| 412 | " \ |
| 413 | ldx %3, %%o0; \ |
| 414 | ldx %4, %%o1; \ |
| 415 | ld %5, %%o2; \ |
| 416 | ld %6, %%o3; \ |
| 417 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 418 | |
| 419 | #define MULADDC_CORE \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 420 | " \ |
| 421 | ld [%%o0], %%o4; \ |
| 422 | inc 4, %%o0; \ |
| 423 | ld [%%o1], %%o5; \ |
| 424 | umul %%o3, %%o4, %%o4; \ |
| 425 | addcc %%o4, %%o2, %%o4; \ |
| 426 | rd %%y, %%g1; \ |
| 427 | addx %%g1, 0, %%g1; \ |
| 428 | addcc %%o4, %%o5, %%o4; \ |
| 429 | st %%o4, [%%o1]; \ |
| 430 | addx %%g1, 0, %%o2; \ |
| 431 | inc 4, %%o1; \ |
| 432 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 433 | |
| 434 | #define MULADDC_STOP \ |
Paul Bakker | 4f024b7 | 2012-10-30 07:29:57 +0000 | [diff] [blame] | 435 | " \ |
| 436 | st %%o2, %0; \ |
| 437 | stx %%o1, %1; \ |
| 438 | stx %%o0, %2; \ |
| 439 | " \ |
| 440 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 441 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 442 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| 443 | "o5" \ |
| 444 | ); |
| 445 | #endif /* SPARCv9 */ |
| 446 | |
| 447 | #if defined(__sparc__) && !defined(__sparc64__) |
| 448 | |
| 449 | #define MULADDC_INIT \ |
| 450 | asm( \ |
| 451 | " \ |
| 452 | ld %3, %%o0; \ |
| 453 | ld %4, %%o1; \ |
| 454 | ld %5, %%o2; \ |
| 455 | ld %6, %%o3; \ |
| 456 | " |
| 457 | |
| 458 | #define MULADDC_CORE \ |
| 459 | " \ |
| 460 | ld [%%o0], %%o4; \ |
| 461 | inc 4, %%o0; \ |
| 462 | ld [%%o1], %%o5; \ |
| 463 | umul %%o3, %%o4, %%o4; \ |
| 464 | addcc %%o4, %%o2, %%o4; \ |
| 465 | rd %%y, %%g1; \ |
| 466 | addx %%g1, 0, %%g1; \ |
| 467 | addcc %%o4, %%o5, %%o4; \ |
| 468 | st %%o4, [%%o1]; \ |
| 469 | addx %%g1, 0, %%o2; \ |
| 470 | inc 4, %%o1; \ |
| 471 | " |
| 472 | |
| 473 | #define MULADDC_STOP \ |
| 474 | " \ |
| 475 | st %%o2, %0; \ |
| 476 | st %%o1, %1; \ |
| 477 | st %%o0, %2; \ |
| 478 | " \ |
| 479 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 480 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 481 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| 482 | "o5" \ |
| 483 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 484 | |
| 485 | #endif /* SPARCv8 */ |
| 486 | |
| 487 | #if defined(__microblaze__) || defined(microblaze) |
| 488 | |
| 489 | #define MULADDC_INIT \ |
| 490 | asm( "lwi r3, %0 " :: "m" (s)); \ |
| 491 | asm( "lwi r4, %0 " :: "m" (d)); \ |
| 492 | asm( "lwi r5, %0 " :: "m" (c)); \ |
| 493 | asm( "lwi r6, %0 " :: "m" (b)); \ |
| 494 | asm( "andi r7, r6, 0xffff" ); \ |
| 495 | asm( "bsrli r6, r6, 16 " ); |
| 496 | |
| 497 | #define MULADDC_CORE \ |
| 498 | asm( "lhui r8, r3, 0 " ); \ |
| 499 | asm( "addi r3, r3, 2 " ); \ |
| 500 | asm( "lhui r9, r3, 0 " ); \ |
| 501 | asm( "addi r3, r3, 2 " ); \ |
| 502 | asm( "mul r10, r9, r6 " ); \ |
| 503 | asm( "mul r11, r8, r7 " ); \ |
| 504 | asm( "mul r12, r9, r7 " ); \ |
| 505 | asm( "mul r13, r8, r6 " ); \ |
| 506 | asm( "bsrli r8, r10, 16 " ); \ |
| 507 | asm( "bsrli r9, r11, 16 " ); \ |
| 508 | asm( "add r13, r13, r8 " ); \ |
| 509 | asm( "add r13, r13, r9 " ); \ |
| 510 | asm( "bslli r10, r10, 16 " ); \ |
| 511 | asm( "bslli r11, r11, 16 " ); \ |
| 512 | asm( "add r12, r12, r10 " ); \ |
| 513 | asm( "addc r13, r13, r0 " ); \ |
| 514 | asm( "add r12, r12, r11 " ); \ |
| 515 | asm( "addc r13, r13, r0 " ); \ |
| 516 | asm( "lwi r10, r4, 0 " ); \ |
| 517 | asm( "add r12, r12, r10 " ); \ |
| 518 | asm( "addc r13, r13, r0 " ); \ |
| 519 | asm( "add r12, r12, r5 " ); \ |
| 520 | asm( "addc r5, r13, r0 " ); \ |
| 521 | asm( "swi r12, r4, 0 " ); \ |
| 522 | asm( "addi r4, r4, 4 " ); |
| 523 | |
| 524 | #define MULADDC_STOP \ |
| 525 | asm( "swi r5, %0 " : "=m" (c)); \ |
| 526 | asm( "swi r4, %0 " : "=m" (d)); \ |
| 527 | asm( "swi r3, %0 " : "=m" (s) :: \ |
| 528 | "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \ |
| 529 | "r9", "r10", "r11", "r12", "r13" ); |
| 530 | |
| 531 | #endif /* MicroBlaze */ |
| 532 | |
| 533 | #if defined(__tricore__) |
| 534 | |
| 535 | #define MULADDC_INIT \ |
| 536 | asm( "ld.a %%a2, %0 " :: "m" (s)); \ |
| 537 | asm( "ld.a %%a3, %0 " :: "m" (d)); \ |
| 538 | asm( "ld.w %%d4, %0 " :: "m" (c)); \ |
| 539 | asm( "ld.w %%d1, %0 " :: "m" (b)); \ |
| 540 | asm( "xor %d5, %d5 " ); |
| 541 | |
| 542 | #define MULADDC_CORE \ |
| 543 | asm( "ld.w %d0, [%a2+] " ); \ |
| 544 | asm( "madd.u %e2, %e4, %d0, %d1 " ); \ |
| 545 | asm( "ld.w %d0, [%a3] " ); \ |
| 546 | asm( "addx %d2, %d2, %d0 " ); \ |
| 547 | asm( "addc %d3, %d3, 0 " ); \ |
| 548 | asm( "mov %d4, %d3 " ); \ |
| 549 | asm( "st.w [%a3+], %d2 " ); |
| 550 | |
| 551 | #define MULADDC_STOP \ |
| 552 | asm( "st.w %0, %%d4 " : "=m" (c)); \ |
| 553 | asm( "st.a %0, %%a3 " : "=m" (d)); \ |
| 554 | asm( "st.a %0, %%a2 " : "=m" (s) :: \ |
| 555 | "d0", "d1", "e2", "d4", "a2", "a3" ); |
| 556 | |
| 557 | #endif /* TriCore */ |
| 558 | |
| 559 | #if defined(__arm__) |
| 560 | |
Paul Bakker | fc4f46f | 2013-06-24 19:23:56 +0200 | [diff] [blame] | 561 | #if defined(__thumb__) && !defined(__thumb2__) |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 562 | |
| 563 | #define MULADDC_INIT \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 564 | asm( \ |
| 565 | " \ |
| 566 | ldr r0, %3; \ |
| 567 | ldr r1, %4; \ |
| 568 | ldr r2, %5; \ |
| 569 | ldr r3, %6; \ |
| 570 | lsr r7, r3, #16; \ |
| 571 | mov r9, r7; \ |
| 572 | lsl r7, r3, #16; \ |
| 573 | lsr r7, r7, #16; \ |
| 574 | mov r8, r7; \ |
| 575 | " |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 576 | |
| 577 | #define MULADDC_CORE \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 578 | " \ |
| 579 | ldmia r0!, {r6}; \ |
| 580 | lsr r7, r6, #16; \ |
| 581 | lsl r6, r6, #16; \ |
| 582 | lsr r6, r6, #16; \ |
| 583 | mov r4, r8; \ |
| 584 | mul r4, r6; \ |
| 585 | mov r3, r9; \ |
| 586 | mul r6, r3; \ |
| 587 | mov r5, r9; \ |
| 588 | mul r5, r7; \ |
| 589 | mov r3, r8; \ |
| 590 | mul r7, r3; \ |
| 591 | lsr r3, r6, #16; \ |
| 592 | add r5, r5, r3; \ |
| 593 | lsr r3, r7, #16; \ |
| 594 | add r5, r5, r3; \ |
| 595 | add r4, r4, r2; \ |
| 596 | mov r2, #0; \ |
| 597 | adc r5, r2; \ |
| 598 | lsl r3, r6, #16; \ |
| 599 | add r4, r4, r3; \ |
| 600 | adc r5, r2; \ |
| 601 | lsl r3, r7, #16; \ |
| 602 | add r4, r4, r3; \ |
| 603 | adc r5, r2; \ |
| 604 | ldr r3, [r1]; \ |
| 605 | add r4, r4, r3; \ |
| 606 | adc r2, r5; \ |
| 607 | stmia r1!, {r4}; \ |
| 608 | " |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 609 | |
| 610 | #define MULADDC_STOP \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 611 | " \ |
| 612 | str r2, %0; \ |
| 613 | str r1, %1; \ |
| 614 | str r0, %2; \ |
| 615 | " \ |
| 616 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 617 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 618 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
Paul Bakker | eff2e6d | 2013-04-11 17:13:22 +0200 | [diff] [blame] | 619 | "r6", "r7", "r8", "r9", "cc" \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 620 | ); |
Paul Bakker | 4f9a7bb | 2012-07-02 08:36:36 +0000 | [diff] [blame] | 621 | |
| 622 | #else |
Paul Bakker | a2713a3 | 2011-11-18 12:47:23 +0000 | [diff] [blame] | 623 | |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 624 | #define MULADDC_INIT \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 625 | asm( \ |
| 626 | " \ |
| 627 | ldr r0, %3; \ |
| 628 | ldr r1, %4; \ |
| 629 | ldr r2, %5; \ |
| 630 | ldr r3, %6; \ |
| 631 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 632 | |
| 633 | #define MULADDC_CORE \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 634 | " \ |
| 635 | ldr r4, [r0], #4; \ |
| 636 | mov r5, #0; \ |
| 637 | ldr r6, [r1]; \ |
| 638 | umlal r2, r5, r3, r4; \ |
| 639 | adds r7, r6, r2; \ |
| 640 | adc r2, r5, #0; \ |
| 641 | str r7, [r1], #4; \ |
| 642 | " |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 643 | |
| 644 | #define MULADDC_STOP \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 645 | " \ |
| 646 | str r2, %0; \ |
| 647 | str r1, %1; \ |
| 648 | str r0, %2; \ |
| 649 | " \ |
| 650 | : "=m" (c), "=m" (d), "=m" (s) \ |
| 651 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| 652 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
Paul Bakker | eff2e6d | 2013-04-11 17:13:22 +0200 | [diff] [blame] | 653 | "r6", "r7", "cc" \ |
Paul Bakker | fb1cbd3 | 2013-03-06 18:14:52 +0100 | [diff] [blame] | 654 | ); |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 655 | |
Paul Bakker | a2713a3 | 2011-11-18 12:47:23 +0000 | [diff] [blame] | 656 | #endif /* Thumb */ |
| 657 | |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 658 | #endif /* ARMv3 */ |
| 659 | |
| 660 | #if defined(__alpha__) |
| 661 | |
| 662 | #define MULADDC_INIT \ |
| 663 | asm( "ldq $1, %0 " :: "m" (s)); \ |
| 664 | asm( "ldq $2, %0 " :: "m" (d)); \ |
| 665 | asm( "ldq $3, %0 " :: "m" (c)); \ |
| 666 | asm( "ldq $4, %0 " :: "m" (b)); |
| 667 | |
| 668 | #define MULADDC_CORE \ |
| 669 | asm( "ldq $6, 0($1) " ); \ |
| 670 | asm( "addq $1, 8, $1 " ); \ |
| 671 | asm( "mulq $6, $4, $7 " ); \ |
| 672 | asm( "umulh $6, $4, $6 " ); \ |
| 673 | asm( "addq $7, $3, $7 " ); \ |
| 674 | asm( "cmpult $7, $3, $3 " ); \ |
| 675 | asm( "ldq $5, 0($2) " ); \ |
| 676 | asm( "addq $7, $5, $7 " ); \ |
| 677 | asm( "cmpult $7, $5, $5 " ); \ |
| 678 | asm( "stq $7, 0($2) " ); \ |
| 679 | asm( "addq $2, 8, $2 " ); \ |
| 680 | asm( "addq $6, $3, $3 " ); \ |
| 681 | asm( "addq $5, $3, $3 " ); |
| 682 | |
| 683 | #define MULADDC_STOP \ |
| 684 | asm( "stq $3, %0 " : "=m" (c)); \ |
| 685 | asm( "stq $2, %0 " : "=m" (d)); \ |
| 686 | asm( "stq $1, %0 " : "=m" (s) :: \ |
| 687 | "$1", "$2", "$3", "$4", "$5", "$6", "$7" ); |
| 688 | |
| 689 | #endif /* Alpha */ |
| 690 | |
| 691 | #if defined(__mips__) |
| 692 | |
| 693 | #define MULADDC_INIT \ |
| 694 | asm( "lw $10, %0 " :: "m" (s)); \ |
| 695 | asm( "lw $11, %0 " :: "m" (d)); \ |
| 696 | asm( "lw $12, %0 " :: "m" (c)); \ |
| 697 | asm( "lw $13, %0 " :: "m" (b)); |
| 698 | |
| 699 | #define MULADDC_CORE \ |
| 700 | asm( "lw $14, 0($10) " ); \ |
| 701 | asm( "multu $13, $14 " ); \ |
| 702 | asm( "addi $10, $10, 4 " ); \ |
| 703 | asm( "mflo $14 " ); \ |
| 704 | asm( "mfhi $9 " ); \ |
| 705 | asm( "addu $14, $12, $14 " ); \ |
| 706 | asm( "lw $15, 0($11) " ); \ |
| 707 | asm( "sltu $12, $14, $12 " ); \ |
| 708 | asm( "addu $15, $14, $15 " ); \ |
| 709 | asm( "sltu $14, $15, $14 " ); \ |
| 710 | asm( "addu $12, $12, $9 " ); \ |
| 711 | asm( "sw $15, 0($11) " ); \ |
| 712 | asm( "addu $12, $12, $14 " ); \ |
| 713 | asm( "addi $11, $11, 4 " ); |
| 714 | |
| 715 | #define MULADDC_STOP \ |
| 716 | asm( "sw $12, %0 " : "=m" (c)); \ |
| 717 | asm( "sw $11, %0 " : "=m" (d)); \ |
| 718 | asm( "sw $10, %0 " : "=m" (s) :: \ |
| 719 | "$9", "$10", "$11", "$12", "$13", "$14", "$15" ); |
| 720 | |
| 721 | #endif /* MIPS */ |
| 722 | #endif /* GNUC */ |
| 723 | |
| 724 | #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) |
| 725 | |
| 726 | #define MULADDC_INIT \ |
| 727 | __asm mov esi, s \ |
| 728 | __asm mov edi, d \ |
| 729 | __asm mov ecx, c \ |
| 730 | __asm mov ebx, b |
| 731 | |
| 732 | #define MULADDC_CORE \ |
| 733 | __asm lodsd \ |
| 734 | __asm mul ebx \ |
| 735 | __asm add eax, ecx \ |
| 736 | __asm adc edx, 0 \ |
| 737 | __asm add eax, [edi] \ |
| 738 | __asm adc edx, 0 \ |
| 739 | __asm mov ecx, edx \ |
| 740 | __asm stosd |
| 741 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 742 | #if defined(POLARSSL_HAVE_SSE2) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 743 | |
| 744 | #define EMIT __asm _emit |
| 745 | |
| 746 | #define MULADDC_HUIT \ |
| 747 | EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ |
| 748 | EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ |
| 749 | EMIT 0x0F EMIT 0x6E EMIT 0x1F \ |
| 750 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 751 | EMIT 0x0F EMIT 0x6E EMIT 0x16 \ |
| 752 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| 753 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ |
| 754 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| 755 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ |
| 756 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| 757 | EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ |
| 758 | EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ |
| 759 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| 760 | EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ |
| 761 | EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ |
| 762 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ |
| 763 | EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ |
| 764 | EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ |
| 765 | EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ |
| 766 | EMIT 0x0F EMIT 0x7E EMIT 0x0F \ |
| 767 | EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ |
| 768 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| 769 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 770 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ |
| 771 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| 772 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 773 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ |
| 774 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| 775 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ |
| 776 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 777 | EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ |
| 778 | EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ |
| 779 | EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ |
| 780 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ |
| 781 | EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ |
| 782 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ |
| 783 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 784 | EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ |
| 785 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ |
| 786 | EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ |
| 787 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ |
| 788 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 789 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| 790 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ |
| 791 | EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ |
| 792 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ |
| 793 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 794 | EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ |
| 795 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ |
| 796 | EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ |
| 797 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ |
| 798 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 799 | EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ |
| 800 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ |
| 801 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 802 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| 803 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ |
| 804 | EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ |
| 805 | EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ |
| 806 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| 807 | EMIT 0x0F EMIT 0x7E EMIT 0xC9 |
| 808 | |
| 809 | #define MULADDC_STOP \ |
| 810 | EMIT 0x0F EMIT 0x77 \ |
| 811 | __asm mov c, ecx \ |
| 812 | __asm mov d, edi \ |
| 813 | __asm mov s, esi \ |
| 814 | |
| 815 | #else |
| 816 | |
| 817 | #define MULADDC_STOP \ |
| 818 | __asm mov c, ecx \ |
| 819 | __asm mov d, edi \ |
| 820 | __asm mov s, esi \ |
| 821 | |
| 822 | #endif /* SSE2 */ |
| 823 | #endif /* MSVC */ |
| 824 | |
Paul Bakker | 40e4694 | 2009-01-03 21:51:57 +0000 | [diff] [blame] | 825 | #endif /* POLARSSL_HAVE_ASM */ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 826 | |
| 827 | #if !defined(MULADDC_CORE) |
Paul Bakker | 0fd018e | 2012-10-23 12:44:47 +0000 | [diff] [blame] | 828 | #if defined(POLARSSL_HAVE_UDBL) |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 829 | |
| 830 | #define MULADDC_INIT \ |
| 831 | { \ |
Paul Bakker | a755ca1 | 2011-04-24 09:11:17 +0000 | [diff] [blame] | 832 | t_udbl r; \ |
| 833 | t_uint r0, r1; |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 834 | |
| 835 | #define MULADDC_CORE \ |
Paul Bakker | 5c2364c | 2012-10-01 14:41:15 +0000 | [diff] [blame] | 836 | r = *(s++) * (t_udbl) b; \ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 837 | r0 = r; \ |
| 838 | r1 = r >> biL; \ |
| 839 | r0 += c; r1 += (r0 < c); \ |
| 840 | r0 += *d; r1 += (r0 < *d); \ |
| 841 | c = r1; *(d++) = r0; |
| 842 | |
| 843 | #define MULADDC_STOP \ |
| 844 | } |
| 845 | |
| 846 | #else |
| 847 | #define MULADDC_INIT \ |
| 848 | { \ |
Paul Bakker | a755ca1 | 2011-04-24 09:11:17 +0000 | [diff] [blame] | 849 | t_uint s0, s1, b0, b1; \ |
| 850 | t_uint r0, r1, rx, ry; \ |
Paul Bakker | 5121ce5 | 2009-01-03 21:22:43 +0000 | [diff] [blame] | 851 | b0 = ( b << biH ) >> biH; \ |
| 852 | b1 = ( b >> biH ); |
| 853 | |
| 854 | #define MULADDC_CORE \ |
| 855 | s0 = ( *s << biH ) >> biH; \ |
| 856 | s1 = ( *s >> biH ); s++; \ |
| 857 | rx = s0 * b1; r0 = s0 * b0; \ |
| 858 | ry = s1 * b0; r1 = s1 * b1; \ |
| 859 | r1 += ( rx >> biH ); \ |
| 860 | r1 += ( ry >> biH ); \ |
| 861 | rx <<= biH; ry <<= biH; \ |
| 862 | r0 += rx; r1 += (r0 < rx); \ |
| 863 | r0 += ry; r1 += (r0 < ry); \ |
| 864 | r0 += c; r1 += (r0 < c); \ |
| 865 | r0 += *d; r1 += (r0 < *d); \ |
| 866 | c = r1; *(d++) = r0; |
| 867 | |
| 868 | #define MULADDC_STOP \ |
| 869 | } |
| 870 | |
| 871 | #endif /* C (generic) */ |
| 872 | #endif /* C (longlong) */ |
| 873 | |
| 874 | #endif /* bn_mul.h */ |