PolarSSL
|
00001 00025 /* 00026 * Multiply source vector [s] with b, add result 00027 * to destination vector [d] and set carry c. 00028 * 00029 * Currently supports: 00030 * 00031 * . IA-32 (386+) . AMD64 / EM64T 00032 * . IA-32 (SSE2) . Motorola 68000 00033 * . PowerPC, 32-bit . MicroBlaze 00034 * . PowerPC, 64-bit . TriCore 00035 * . SPARC v8 . ARM v3+ 00036 * . Alpha . MIPS32 00037 * . C, longlong . C, generic 00038 */ 00039 #ifndef POLARSSL_BN_MUL_H 00040 #define POLARSSL_BN_MUL_H 00041 00042 #include "polarssl/config.h" 00043 00044 #if defined(POLARSSL_HAVE_ASM) 00045 00046 #if defined(__GNUC__) 00047 #if defined(__i386__) 00048 00049 #define MULADDC_INIT \ 00050 asm( " \ 00051 movl %%ebx, %0; \ 00052 movl %5, %%esi; \ 00053 movl %6, %%edi; \ 00054 movl %7, %%ecx; \ 00055 movl %8, %%ebx; \ 00056 " 00057 00058 #define MULADDC_CORE \ 00059 " \ 00060 lodsl; \ 00061 mull %%ebx; \ 00062 addl %%ecx, %%eax; \ 00063 adcl $0, %%edx; \ 00064 addl (%%edi), %%eax; \ 00065 adcl $0, %%edx; \ 00066 movl %%edx, %%ecx; \ 00067 stosl; \ 00068 " 00069 00070 #if defined(POLARSSL_HAVE_SSE2) 00071 00072 #define MULADDC_HUIT \ 00073 " \ 00074 movd %%ecx, %%mm1; \ 00075 movd %%ebx, %%mm0; \ 00076 movd (%%edi), %%mm3; \ 00077 paddq %%mm3, %%mm1; \ 00078 movd (%%esi), %%mm2; \ 00079 pmuludq %%mm0, %%mm2; \ 00080 movd 4(%%esi), %%mm4; \ 00081 pmuludq %%mm0, %%mm4; \ 00082 movd 8(%%esi), %%mm6; \ 00083 pmuludq %%mm0, %%mm6; \ 00084 movd 12(%%esi), %%mm7; \ 00085 pmuludq %%mm0, %%mm7; \ 00086 paddq %%mm2, %%mm1; \ 00087 movd 4(%%edi), %%mm3; \ 00088 paddq %%mm4, %%mm3; \ 00089 movd 8(%%edi), %%mm5; \ 00090 paddq %%mm6, %%mm5; \ 00091 movd 12(%%edi), %%mm4; \ 00092 paddq %%mm4, %%mm7; \ 00093 movd %%mm1, (%%edi); \ 00094 movd 16(%%esi), %%mm2; \ 00095 pmuludq %%mm0, %%mm2; \ 00096 psrlq $32, %%mm1; \ 00097 movd 20(%%esi), %%mm4; \ 00098 pmuludq %%mm0, %%mm4; \ 00099 paddq %%mm3, %%mm1; \ 00100 movd 24(%%esi), %%mm6; \ 00101 pmuludq %%mm0, %%mm6; \ 00102 movd %%mm1, 4(%%edi); \ 00103 psrlq $32, %%mm1; \ 00104 movd 28(%%esi), %%mm3; \ 00105 pmuludq %%mm0, %%mm3; \ 00106 paddq %%mm5, %%mm1; \ 00107 movd 16(%%edi), %%mm5; \ 00108 paddq %%mm5, %%mm2; \ 00109 movd %%mm1, 8(%%edi); \ 00110 psrlq $32, %%mm1; \ 00111 paddq %%mm7, %%mm1; \ 00112 movd 20(%%edi), %%mm5; \ 00113 paddq %%mm5, %%mm4; \ 00114 movd %%mm1, 12(%%edi); \ 00115 psrlq $32, %%mm1; \ 00116 paddq %%mm2, %%mm1; \ 00117 movd 24(%%edi), %%mm5; \ 00118 paddq %%mm5, %%mm6; \ 00119 movd %%mm1, 16(%%edi); \ 00120 psrlq $32, %%mm1; \ 00121 paddq %%mm4, %%mm1; \ 00122 movd 28(%%edi), %%mm5; \ 00123 paddq %%mm5, %%mm3; \ 00124 movd %%mm1, 20(%%edi); \ 00125 psrlq $32, %%mm1; \ 00126 paddq %%mm6, %%mm1; \ 00127 movd %%mm1, 24(%%edi); \ 00128 psrlq $32, %%mm1; \ 00129 paddq %%mm3, %%mm1; \ 00130 movd %%mm1, 28(%%edi); \ 00131 addl $32, %%edi; \ 00132 addl $32, %%esi; \ 00133 psrlq $32, %%mm1; \ 00134 movd %%mm1, %%ecx; \ 00135 " 00136 00137 #define MULADDC_STOP \ 00138 " \ 00139 emms; \ 00140 movl %4, %%ebx; \ 00141 movl %%ecx, %1; \ 00142 movl %%edi, %2; \ 00143 movl %%esi, %3; \ 00144 " \ 00145 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 00146 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 00147 : "eax", "ecx", "edx", "esi", "edi" \ 00148 ); 00149 00150 #else 00151 00152 #define MULADDC_STOP \ 00153 " \ 00154 movl %4, %%ebx; \ 00155 movl %%ecx, %1; \ 00156 movl %%edi, %2; \ 00157 movl %%esi, %3; \ 00158 " \ 00159 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ 00160 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ 00161 : "eax", "ecx", "edx", "esi", "edi" \ 00162 ); 00163 #endif /* SSE2 */ 00164 #endif /* i386 */ 00165 00166 #if defined(__amd64__) || defined (__x86_64__) 00167 00168 #define MULADDC_INIT \ 00169 asm( "movq %0, %%rsi " :: "m" (s)); \ 00170 asm( "movq %0, %%rdi " :: "m" (d)); \ 00171 asm( "movq %0, %%rcx " :: "m" (c)); \ 00172 asm( "movq %0, %%rbx " :: "m" (b)); \ 00173 asm( "xorq %r8, %r8 " ); 00174 00175 #define MULADDC_CORE \ 00176 asm( "movq (%rsi),%rax " ); \ 00177 asm( "mulq %rbx " ); \ 00178 asm( "addq $8, %rsi " ); \ 00179 asm( "addq %rcx, %rax " ); \ 00180 asm( "movq %r8, %rcx " ); \ 00181 asm( "adcq $0, %rdx " ); \ 00182 asm( "nop " ); \ 00183 asm( "addq %rax, (%rdi) " ); \ 00184 asm( "adcq %rdx, %rcx " ); \ 00185 asm( "addq $8, %rdi " ); 00186 00187 #define MULADDC_STOP \ 00188 asm( "movq %%rcx, %0 " : "=m" (c)); \ 00189 asm( "movq %%rdi, %0 " : "=m" (d)); \ 00190 asm( "movq %%rsi, %0 " : "=m" (s) :: \ 00191 "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" ); 00192 00193 #endif /* AMD64 */ 00194 00195 #if defined(__mc68020__) || defined(__mcpu32__) 00196 00197 #define MULADDC_INIT \ 00198 asm( "movl %0, %%a2 " :: "m" (s)); \ 00199 asm( "movl %0, %%a3 " :: "m" (d)); \ 00200 asm( "movl %0, %%d3 " :: "m" (c)); \ 00201 asm( "movl %0, %%d2 " :: "m" (b)); \ 00202 asm( "moveq #0, %d0 " ); 00203 00204 #define MULADDC_CORE \ 00205 asm( "movel %a2@+, %d1 " ); \ 00206 asm( "mulul %d2, %d4:%d1 " ); \ 00207 asm( "addl %d3, %d1 " ); \ 00208 asm( "addxl %d0, %d4 " ); \ 00209 asm( "moveq #0, %d3 " ); \ 00210 asm( "addl %d1, %a3@+ " ); \ 00211 asm( "addxl %d4, %d3 " ); 00212 00213 #define MULADDC_STOP \ 00214 asm( "movl %%d3, %0 " : "=m" (c)); \ 00215 asm( "movl %%a3, %0 " : "=m" (d)); \ 00216 asm( "movl %%a2, %0 " : "=m" (s) :: \ 00217 "d0", "d1", "d2", "d3", "d4", "a2", "a3" ); 00218 00219 #define MULADDC_HUIT \ 00220 asm( "movel %a2@+, %d1 " ); \ 00221 asm( "mulul %d2, %d4:%d1 " ); \ 00222 asm( "addxl %d3, %d1 " ); \ 00223 asm( "addxl %d0, %d4 " ); \ 00224 asm( "addl %d1, %a3@+ " ); \ 00225 asm( "movel %a2@+, %d1 " ); \ 00226 asm( "mulul %d2, %d3:%d1 " ); \ 00227 asm( "addxl %d4, %d1 " ); \ 00228 asm( "addxl %d0, %d3 " ); \ 00229 asm( "addl %d1, %a3@+ " ); \ 00230 asm( "movel %a2@+, %d1 " ); \ 00231 asm( "mulul %d2, %d4:%d1 " ); \ 00232 asm( "addxl %d3, %d1 " ); \ 00233 asm( "addxl %d0, %d4 " ); \ 00234 asm( "addl %d1, %a3@+ " ); \ 00235 asm( "movel %a2@+, %d1 " ); \ 00236 asm( "mulul %d2, %d3:%d1 " ); \ 00237 asm( "addxl %d4, %d1 " ); \ 00238 asm( "addxl %d0, %d3 " ); \ 00239 asm( "addl %d1, %a3@+ " ); \ 00240 asm( "movel %a2@+, %d1 " ); \ 00241 asm( "mulul %d2, %d4:%d1 " ); \ 00242 asm( "addxl %d3, %d1 " ); \ 00243 asm( "addxl %d0, %d4 " ); \ 00244 asm( "addl %d1, %a3@+ " ); \ 00245 asm( "movel %a2@+, %d1 " ); \ 00246 asm( "mulul %d2, %d3:%d1 " ); \ 00247 asm( "addxl %d4, %d1 " ); \ 00248 asm( "addxl %d0, %d3 " ); \ 00249 asm( "addl %d1, %a3@+ " ); \ 00250 asm( "movel %a2@+, %d1 " ); \ 00251 asm( "mulul %d2, %d4:%d1 " ); \ 00252 asm( "addxl %d3, %d1 " ); \ 00253 asm( "addxl %d0, %d4 " ); \ 00254 asm( "addl %d1, %a3@+ " ); \ 00255 asm( "movel %a2@+, %d1 " ); \ 00256 asm( "mulul %d2, %d3:%d1 " ); \ 00257 asm( "addxl %d4, %d1 " ); \ 00258 asm( "addxl %d0, %d3 " ); \ 00259 asm( "addl %d1, %a3@+ " ); \ 00260 asm( "addxl %d0, %d3 " ); 00261 00262 #endif /* MC68000 */ 00263 00264 #if defined(__powerpc__) || defined(__ppc__) 00265 #if defined(__powerpc64__) || defined(__ppc64__) 00266 00267 #if defined(__MACH__) && defined(__APPLE__) 00268 00269 #define MULADDC_INIT \ 00270 asm( "ld r3, %0 " :: "m" (s)); \ 00271 asm( "ld r4, %0 " :: "m" (d)); \ 00272 asm( "ld r5, %0 " :: "m" (c)); \ 00273 asm( "ld r6, %0 " :: "m" (b)); \ 00274 asm( "addi r3, r3, -8 " ); \ 00275 asm( "addi r4, r4, -8 " ); \ 00276 asm( "addic r5, r5, 0 " ); 00277 00278 #define MULADDC_CORE \ 00279 asm( "ldu r7, 8(r3) " ); \ 00280 asm( "mulld r8, r7, r6 " ); \ 00281 asm( "mulhdu r9, r7, r6 " ); \ 00282 asm( "adde r8, r8, r5 " ); \ 00283 asm( "ld r7, 8(r4) " ); \ 00284 asm( "addze r5, r9 " ); \ 00285 asm( "addc r8, r8, r7 " ); \ 00286 asm( "stdu r8, 8(r4) " ); 00287 00288 #define MULADDC_STOP \ 00289 asm( "addze r5, r5 " ); \ 00290 asm( "addi r4, r4, 8 " ); \ 00291 asm( "addi r3, r3, 8 " ); \ 00292 asm( "std r5, %0 " : "=m" (c)); \ 00293 asm( "std r4, %0 " : "=m" (d)); \ 00294 asm( "std r3, %0 " : "=m" (s) :: \ 00295 "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); 00296 00297 #else 00298 00299 #define MULADDC_INIT \ 00300 asm( "ld %%r3, %0 " :: "m" (s)); \ 00301 asm( "ld %%r4, %0 " :: "m" (d)); \ 00302 asm( "ld %%r5, %0 " :: "m" (c)); \ 00303 asm( "ld %%r6, %0 " :: "m" (b)); \ 00304 asm( "addi %r3, %r3, -8 " ); \ 00305 asm( "addi %r4, %r4, -8 " ); \ 00306 asm( "addic %r5, %r5, 0 " ); 00307 00308 #define MULADDC_CORE \ 00309 asm( "ldu %r7, 8(%r3) " ); \ 00310 asm( "mulld %r8, %r7, %r6 " ); \ 00311 asm( "mulhdu %r9, %r7, %r6 " ); \ 00312 asm( "adde %r8, %r8, %r5 " ); \ 00313 asm( "ld %r7, 8(%r4) " ); \ 00314 asm( "addze %r5, %r9 " ); \ 00315 asm( "addc %r8, %r8, %r7 " ); \ 00316 asm( "stdu %r8, 8(%r4) " ); 00317 00318 #define MULADDC_STOP \ 00319 asm( "addze %r5, %r5 " ); \ 00320 asm( "addi %r4, %r4, 8 " ); \ 00321 asm( "addi %r3, %r3, 8 " ); \ 00322 asm( "std %%r5, %0 " : "=m" (c)); \ 00323 asm( "std %%r4, %0 " : "=m" (d)); \ 00324 asm( "std %%r3, %0 " : "=m" (s) :: \ 00325 "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); 00326 00327 #endif 00328 00329 #else /* PPC32 */ 00330 00331 #if defined(__MACH__) && defined(__APPLE__) 00332 00333 #define MULADDC_INIT \ 00334 asm( "lwz r3, %0 " :: "m" (s)); \ 00335 asm( "lwz r4, %0 " :: "m" (d)); \ 00336 asm( "lwz r5, %0 " :: "m" (c)); \ 00337 asm( "lwz r6, %0 " :: "m" (b)); \ 00338 asm( "addi r3, r3, -4 " ); \ 00339 asm( "addi r4, r4, -4 " ); \ 00340 asm( "addic r5, r5, 0 " ); 00341 00342 #define MULADDC_CORE \ 00343 asm( "lwzu r7, 4(r3) " ); \ 00344 asm( "mullw r8, r7, r6 " ); \ 00345 asm( "mulhwu r9, r7, r6 " ); \ 00346 asm( "adde r8, r8, r5 " ); \ 00347 asm( "lwz r7, 4(r4) " ); \ 00348 asm( "addze r5, r9 " ); \ 00349 asm( "addc r8, r8, r7 " ); \ 00350 asm( "stwu r8, 4(r4) " ); 00351 00352 #define MULADDC_STOP \ 00353 asm( "addze r5, r5 " ); \ 00354 asm( "addi r4, r4, 4 " ); \ 00355 asm( "addi r3, r3, 4 " ); \ 00356 asm( "stw r5, %0 " : "=m" (c)); \ 00357 asm( "stw r4, %0 " : "=m" (d)); \ 00358 asm( "stw r3, %0 " : "=m" (s) :: \ 00359 "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); 00360 00361 #else 00362 00363 #define MULADDC_INIT \ 00364 asm( "lwz %%r3, %0 " :: "m" (s)); \ 00365 asm( "lwz %%r4, %0 " :: "m" (d)); \ 00366 asm( "lwz %%r5, %0 " :: "m" (c)); \ 00367 asm( "lwz %%r6, %0 " :: "m" (b)); \ 00368 asm( "addi %r3, %r3, -4 " ); \ 00369 asm( "addi %r4, %r4, -4 " ); \ 00370 asm( "addic %r5, %r5, 0 " ); 00371 00372 #define MULADDC_CORE \ 00373 asm( "lwzu %r7, 4(%r3) " ); \ 00374 asm( "mullw %r8, %r7, %r6 " ); \ 00375 asm( "mulhwu %r9, %r7, %r6 " ); \ 00376 asm( "adde %r8, %r8, %r5 " ); \ 00377 asm( "lwz %r7, 4(%r4) " ); \ 00378 asm( "addze %r5, %r9 " ); \ 00379 asm( "addc %r8, %r8, %r7 " ); \ 00380 asm( "stwu %r8, 4(%r4) " ); 00381 00382 #define MULADDC_STOP \ 00383 asm( "addze %r5, %r5 " ); \ 00384 asm( "addi %r4, %r4, 4 " ); \ 00385 asm( "addi %r3, %r3, 4 " ); \ 00386 asm( "stw %%r5, %0 " : "=m" (c)); \ 00387 asm( "stw %%r4, %0 " : "=m" (d)); \ 00388 asm( "stw %%r3, %0 " : "=m" (s) :: \ 00389 "r3", "r4", "r5", "r6", "r7", "r8", "r9" ); 00390 00391 #endif 00392 00393 #endif /* PPC32 */ 00394 #endif /* PPC64 */ 00395 00396 #if defined(__sparc__) 00397 00398 #define MULADDC_INIT \ 00399 asm( "ld %0, %%o0 " :: "m" (s)); \ 00400 asm( "ld %0, %%o1 " :: "m" (d)); \ 00401 asm( "ld %0, %%o2 " :: "m" (c)); \ 00402 asm( "ld %0, %%o3 " :: "m" (b)); 00403 00404 #define MULADDC_CORE \ 00405 asm( "ld [%o0], %o4 " ); \ 00406 asm( "inc 4, %o0 " ); \ 00407 asm( "ld [%o1], %o5 " ); \ 00408 asm( "umul %o3, %o4, %o4 " ); \ 00409 asm( "addcc %o4, %o2, %o4 " ); \ 00410 asm( "rd %y, %g1 " ); \ 00411 asm( "addx %g1, 0, %g1 " ); \ 00412 asm( "addcc %o4, %o5, %o4 " ); \ 00413 asm( "st %o4, [%o1] " ); \ 00414 asm( "addx %g1, 0, %o2 " ); \ 00415 asm( "inc 4, %o1 " ); 00416 00417 #define MULADDC_STOP \ 00418 asm( "st %%o2, %0 " : "=m" (c)); \ 00419 asm( "st %%o1, %0 " : "=m" (d)); \ 00420 asm( "st %%o0, %0 " : "=m" (s) :: \ 00421 "g1", "o0", "o1", "o2", "o3", "o4", "o5" ); 00422 00423 #endif /* SPARCv8 */ 00424 00425 #if defined(__microblaze__) || defined(microblaze) 00426 00427 #define MULADDC_INIT \ 00428 asm( "lwi r3, %0 " :: "m" (s)); \ 00429 asm( "lwi r4, %0 " :: "m" (d)); \ 00430 asm( "lwi r5, %0 " :: "m" (c)); \ 00431 asm( "lwi r6, %0 " :: "m" (b)); \ 00432 asm( "andi r7, r6, 0xffff" ); \ 00433 asm( "bsrli r6, r6, 16 " ); 00434 00435 #define MULADDC_CORE \ 00436 asm( "lhui r8, r3, 0 " ); \ 00437 asm( "addi r3, r3, 2 " ); \ 00438 asm( "lhui r9, r3, 0 " ); \ 00439 asm( "addi r3, r3, 2 " ); \ 00440 asm( "mul r10, r9, r6 " ); \ 00441 asm( "mul r11, r8, r7 " ); \ 00442 asm( "mul r12, r9, r7 " ); \ 00443 asm( "mul r13, r8, r6 " ); \ 00444 asm( "bsrli r8, r10, 16 " ); \ 00445 asm( "bsrli r9, r11, 16 " ); \ 00446 asm( "add r13, r13, r8 " ); \ 00447 asm( "add r13, r13, r9 " ); \ 00448 asm( "bslli r10, r10, 16 " ); \ 00449 asm( "bslli r11, r11, 16 " ); \ 00450 asm( "add r12, r12, r10 " ); \ 00451 asm( "addc r13, r13, r0 " ); \ 00452 asm( "add r12, r12, r11 " ); \ 00453 asm( "addc r13, r13, r0 " ); \ 00454 asm( "lwi r10, r4, 0 " ); \ 00455 asm( "add r12, r12, r10 " ); \ 00456 asm( "addc r13, r13, r0 " ); \ 00457 asm( "add r12, r12, r5 " ); \ 00458 asm( "addc r5, r13, r0 " ); \ 00459 asm( "swi r12, r4, 0 " ); \ 00460 asm( "addi r4, r4, 4 " ); 00461 00462 #define MULADDC_STOP \ 00463 asm( "swi r5, %0 " : "=m" (c)); \ 00464 asm( "swi r4, %0 " : "=m" (d)); \ 00465 asm( "swi r3, %0 " : "=m" (s) :: \ 00466 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \ 00467 "r9", "r10", "r11", "r12", "r13" ); 00468 00469 #endif /* MicroBlaze */ 00470 00471 #if defined(__tricore__) 00472 00473 #define MULADDC_INIT \ 00474 asm( "ld.a %%a2, %0 " :: "m" (s)); \ 00475 asm( "ld.a %%a3, %0 " :: "m" (d)); \ 00476 asm( "ld.w %%d4, %0 " :: "m" (c)); \ 00477 asm( "ld.w %%d1, %0 " :: "m" (b)); \ 00478 asm( "xor %d5, %d5 " ); 00479 00480 #define MULADDC_CORE \ 00481 asm( "ld.w %d0, [%a2+] " ); \ 00482 asm( "madd.u %e2, %e4, %d0, %d1 " ); \ 00483 asm( "ld.w %d0, [%a3] " ); \ 00484 asm( "addx %d2, %d2, %d0 " ); \ 00485 asm( "addc %d3, %d3, 0 " ); \ 00486 asm( "mov %d4, %d3 " ); \ 00487 asm( "st.w [%a3+], %d2 " ); 00488 00489 #define MULADDC_STOP \ 00490 asm( "st.w %0, %%d4 " : "=m" (c)); \ 00491 asm( "st.a %0, %%a3 " : "=m" (d)); \ 00492 asm( "st.a %0, %%a2 " : "=m" (s) :: \ 00493 "d0", "d1", "e2", "d4", "a2", "a3" ); 00494 00495 #endif /* TriCore */ 00496 00497 #if defined(__arm__) 00498 00499 #define MULADDC_INIT \ 00500 asm( "ldr r0, %0 " :: "m" (s)); \ 00501 asm( "ldr r1, %0 " :: "m" (d)); \ 00502 asm( "ldr r2, %0 " :: "m" (c)); \ 00503 asm( "ldr r3, %0 " :: "m" (b)); 00504 00505 #define MULADDC_CORE \ 00506 asm( "ldr r4, [r0], #4 " ); \ 00507 asm( "mov r5, #0 " ); \ 00508 asm( "ldr r6, [r1] " ); \ 00509 asm( "umlal r2, r5, r3, r4 " ); \ 00510 asm( "adds r7, r6, r2 " ); \ 00511 asm( "adc r2, r5, #0 " ); \ 00512 asm( "str r7, [r1], #4 " ); 00513 00514 #define MULADDC_STOP \ 00515 asm( "str r2, %0 " : "=m" (c)); \ 00516 asm( "str r1, %0 " : "=m" (d)); \ 00517 asm( "str r0, %0 " : "=m" (s) :: \ 00518 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7" ); 00519 00520 #endif /* ARMv3 */ 00521 00522 #if defined(__alpha__) 00523 00524 #define MULADDC_INIT \ 00525 asm( "ldq $1, %0 " :: "m" (s)); \ 00526 asm( "ldq $2, %0 " :: "m" (d)); \ 00527 asm( "ldq $3, %0 " :: "m" (c)); \ 00528 asm( "ldq $4, %0 " :: "m" (b)); 00529 00530 #define MULADDC_CORE \ 00531 asm( "ldq $6, 0($1) " ); \ 00532 asm( "addq $1, 8, $1 " ); \ 00533 asm( "mulq $6, $4, $7 " ); \ 00534 asm( "umulh $6, $4, $6 " ); \ 00535 asm( "addq $7, $3, $7 " ); \ 00536 asm( "cmpult $7, $3, $3 " ); \ 00537 asm( "ldq $5, 0($2) " ); \ 00538 asm( "addq $7, $5, $7 " ); \ 00539 asm( "cmpult $7, $5, $5 " ); \ 00540 asm( "stq $7, 0($2) " ); \ 00541 asm( "addq $2, 8, $2 " ); \ 00542 asm( "addq $6, $3, $3 " ); \ 00543 asm( "addq $5, $3, $3 " ); 00544 00545 #define MULADDC_STOP \ 00546 asm( "stq $3, %0 " : "=m" (c)); \ 00547 asm( "stq $2, %0 " : "=m" (d)); \ 00548 asm( "stq $1, %0 " : "=m" (s) :: \ 00549 "$1", "$2", "$3", "$4", "$5", "$6", "$7" ); 00550 00551 #endif /* Alpha */ 00552 00553 #if defined(__mips__) 00554 00555 #define MULADDC_INIT \ 00556 asm( "lw $10, %0 " :: "m" (s)); \ 00557 asm( "lw $11, %0 " :: "m" (d)); \ 00558 asm( "lw $12, %0 " :: "m" (c)); \ 00559 asm( "lw $13, %0 " :: "m" (b)); 00560 00561 #define MULADDC_CORE \ 00562 asm( "lw $14, 0($10) " ); \ 00563 asm( "multu $13, $14 " ); \ 00564 asm( "addi $10, $10, 4 " ); \ 00565 asm( "mflo $14 " ); \ 00566 asm( "mfhi $9 " ); \ 00567 asm( "addu $14, $12, $14 " ); \ 00568 asm( "lw $15, 0($11) " ); \ 00569 asm( "sltu $12, $14, $12 " ); \ 00570 asm( "addu $15, $14, $15 " ); \ 00571 asm( "sltu $14, $15, $14 " ); \ 00572 asm( "addu $12, $12, $9 " ); \ 00573 asm( "sw $15, 0($11) " ); \ 00574 asm( "addu $12, $12, $14 " ); \ 00575 asm( "addi $11, $11, 4 " ); 00576 00577 #define MULADDC_STOP \ 00578 asm( "sw $12, %0 " : "=m" (c)); \ 00579 asm( "sw $11, %0 " : "=m" (d)); \ 00580 asm( "sw $10, %0 " : "=m" (s) :: \ 00581 "$9", "$10", "$11", "$12", "$13", "$14", "$15" ); 00582 00583 #endif /* MIPS */ 00584 #endif /* GNUC */ 00585 00586 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) 00587 00588 #define MULADDC_INIT \ 00589 __asm mov esi, s \ 00590 __asm mov edi, d \ 00591 __asm mov ecx, c \ 00592 __asm mov ebx, b 00593 00594 #define MULADDC_CORE \ 00595 __asm lodsd \ 00596 __asm mul ebx \ 00597 __asm add eax, ecx \ 00598 __asm adc edx, 0 \ 00599 __asm add eax, [edi] \ 00600 __asm adc edx, 0 \ 00601 __asm mov ecx, edx \ 00602 __asm stosd 00603 00604 #if defined(POLARSSL_HAVE_SSE2) 00605 00606 #define EMIT __asm _emit 00607 00608 #define MULADDC_HUIT \ 00609 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ 00610 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ 00611 EMIT 0x0F EMIT 0x6E EMIT 0x1F \ 00612 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00613 EMIT 0x0F EMIT 0x6E EMIT 0x16 \ 00614 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 00615 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ 00616 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 00617 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ 00618 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 00619 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ 00620 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ 00621 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 00622 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ 00623 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ 00624 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ 00625 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ 00626 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ 00627 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ 00628 EMIT 0x0F EMIT 0x7E EMIT 0x0F \ 00629 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ 00630 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ 00631 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00632 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ 00633 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ 00634 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00635 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ 00636 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ 00637 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ 00638 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00639 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ 00640 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ 00641 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ 00642 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ 00643 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ 00644 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ 00645 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00646 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ 00647 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ 00648 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ 00649 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ 00650 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00651 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ 00652 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ 00653 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ 00654 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ 00655 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00656 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ 00657 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ 00658 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ 00659 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ 00660 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00661 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ 00662 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ 00663 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00664 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ 00665 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ 00666 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ 00667 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ 00668 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ 00669 EMIT 0x0F EMIT 0x7E EMIT 0xC9 00670 00671 #define MULADDC_STOP \ 00672 EMIT 0x0F EMIT 0x77 \ 00673 __asm mov c, ecx \ 00674 __asm mov d, edi \ 00675 __asm mov s, esi \ 00676 00677 #else 00678 00679 #define MULADDC_STOP \ 00680 __asm mov c, ecx \ 00681 __asm mov d, edi \ 00682 __asm mov s, esi \ 00683 00684 #endif /* SSE2 */ 00685 #endif /* MSVC */ 00686 00687 #endif /* POLARSSL_HAVE_ASM */ 00688 00689 #if !defined(MULADDC_CORE) 00690 #if defined(POLARSSL_HAVE_LONGLONG) 00691 00692 #define MULADDC_INIT \ 00693 { \ 00694 t_dbl r; \ 00695 t_int r0, r1; 00696 00697 #define MULADDC_CORE \ 00698 r = *(s++) * (t_dbl) b; \ 00699 r0 = r; \ 00700 r1 = r >> biL; \ 00701 r0 += c; r1 += (r0 < c); \ 00702 r0 += *d; r1 += (r0 < *d); \ 00703 c = r1; *(d++) = r0; 00704 00705 #define MULADDC_STOP \ 00706 } 00707 00708 #else 00709 #define MULADDC_INIT \ 00710 { \ 00711 t_int s0, s1, b0, b1; \ 00712 t_int r0, r1, rx, ry; \ 00713 b0 = ( b << biH ) >> biH; \ 00714 b1 = ( b >> biH ); 00715 00716 #define MULADDC_CORE \ 00717 s0 = ( *s << biH ) >> biH; \ 00718 s1 = ( *s >> biH ); s++; \ 00719 rx = s0 * b1; r0 = s0 * b0; \ 00720 ry = s1 * b0; r1 = s1 * b1; \ 00721 r1 += ( rx >> biH ); \ 00722 r1 += ( ry >> biH ); \ 00723 rx <<= biH; ry <<= biH; \ 00724 r0 += rx; r1 += (r0 < rx); \ 00725 r0 += ry; r1 += (r0 < ry); \ 00726 r0 += c; r1 += (r0 < c); \ 00727 r0 += *d; r1 += (r0 < *d); \ 00728 c = r1; *(d++) = r0; 00729 00730 #define MULADDC_STOP \ 00731 } 00732 00733 #endif /* C (generic) */ 00734 #endif /* C (longlong) */ 00735 00736 #endif /* bn_mul.h */