00001
00024 #ifndef _BN_ASM_H
00025 #define _BN_ASM_H
00026
00027 #if defined(__GNUC__)
00028 #if defined(__i386__)
00029
00030 #define MULADDC_INIT \
00031 asm( "movl %%ebx, %0 " : "=m" (t)); \
00032 asm( "movl %0, %%esi " :: "m" (s)); \
00033 asm( "movl %0, %%edi " :: "m" (d)); \
00034 asm( "movl %0, %%ecx " :: "m" (c)); \
00035 asm( "movl %0, %%ebx " :: "m" (b));
00036
00037 #define MULADDC_CORE \
00038 asm( "lodsl " ); \
00039 asm( "mull %ebx " ); \
00040 asm( "addl %ecx, %eax " ); \
00041 asm( "adcl $0, %edx " ); \
00042 asm( "addl (%edi), %eax " ); \
00043 asm( "adcl $0, %edx " ); \
00044 asm( "movl %edx, %ecx " ); \
00045 asm( "stosl " );
00046
00047 #define MULADDC_STOP \
00048 asm( "movl %%ecx, %0 " : "=m" (c)); \
00049 asm( "movl %%edi, %0 " : "=m" (d)); \
00050 asm( "movl %%esi, %0 " : "=m" (s)); \
00051 asm( "movl %0, %%ebx " :: "m" (t) : \
00052 "eax", "ecx", "edx", "esi", "edi" );
00053
00054 #if defined(HAVE_SSE2)
00055
00056 #define MULADDC_HUIT \
00057 asm( "movd %ecx, %mm1 " ); \
00058 asm( "movd %ebx, %mm0 " ); \
00059 asm( "movd (%edi), %mm3 " ); \
00060 asm( "paddq %mm3, %mm1 " ); \
00061 asm( "movd (%esi), %mm2 " ); \
00062 asm( "pmuludq %mm0, %mm2 " ); \
00063 asm( "movd 4(%esi), %mm4 " ); \
00064 asm( "pmuludq %mm0, %mm4 " ); \
00065 asm( "movd 8(%esi), %mm6 " ); \
00066 asm( "pmuludq %mm0, %mm6 " ); \
00067 asm( "movd 12(%esi), %mm7 " ); \
00068 asm( "pmuludq %mm0, %mm7 " ); \
00069 asm( "paddq %mm2, %mm1 " ); \
00070 asm( "movd 4(%edi), %mm3 " ); \
00071 asm( "paddq %mm4, %mm3 " ); \
00072 asm( "movd 8(%edi), %mm5 " ); \
00073 asm( "paddq %mm6, %mm5 " ); \
00074 asm( "movd 12(%edi), %mm4 " ); \
00075 asm( "paddq %mm4, %mm7 " ); \
00076 asm( "movd %mm1, (%edi) " ); \
00077 asm( "movd 16(%esi), %mm2 " ); \
00078 asm( "pmuludq %mm0, %mm2 " ); \
00079 asm( "psrlq $32, %mm1 " ); \
00080 asm( "movd 20(%esi), %mm4 " ); \
00081 asm( "pmuludq %mm0, %mm4 " ); \
00082 asm( "paddq %mm3, %mm1 " ); \
00083 asm( "movd 24(%esi), %mm6 " ); \
00084 asm( "pmuludq %mm0, %mm6 " ); \
00085 asm( "movd %mm1, 4(%edi) " ); \
00086 asm( "psrlq $32, %mm1 " ); \
00087 asm( "movd 28(%esi), %mm3 " ); \
00088 asm( "pmuludq %mm0, %mm3 " ); \
00089 asm( "paddq %mm5, %mm1 " ); \
00090 asm( "movd 16(%edi), %mm5 " ); \
00091 asm( "paddq %mm5, %mm2 " ); \
00092 asm( "movd %mm1, 8(%edi) " ); \
00093 asm( "psrlq $32, %mm1 " ); \
00094 asm( "paddq %mm7, %mm1 " ); \
00095 asm( "movd 20(%edi), %mm5 " ); \
00096 asm( "paddq %mm5, %mm4 " ); \
00097 asm( "movd %mm1, 12(%edi) " ); \
00098 asm( "psrlq $32, %mm1 " ); \
00099 asm( "paddq %mm2, %mm1 " ); \
00100 asm( "movd 24(%edi), %mm5 " ); \
00101 asm( "paddq %mm5, %mm6 " ); \
00102 asm( "movd %mm1, 16(%edi) " ); \
00103 asm( "psrlq $32, %mm1 " ); \
00104 asm( "paddq %mm4, %mm1 " ); \
00105 asm( "movd 28(%edi), %mm5 " ); \
00106 asm( "paddq %mm5, %mm3 " ); \
00107 asm( "movd %mm1, 20(%edi) " ); \
00108 asm( "psrlq $32, %mm1 " ); \
00109 asm( "paddq %mm6, %mm1 " ); \
00110 asm( "movd %mm1, 24(%edi) " ); \
00111 asm( "psrlq $32, %mm1 " ); \
00112 asm( "paddq %mm3, %mm1 " ); \
00113 asm( "movd %mm1, 28(%edi) " ); \
00114 asm( "addl $32, %edi " ); \
00115 asm( "addl $32, %esi " ); \
00116 asm( "psrlq $32, %mm1 " ); \
00117 asm( "movd %mm1, %ecx " );
00118
00119 #endif
00120 #endif
00121
00122 #if defined(__amd64__) || defined (__x86_64__)
00123
00124 #define MULADDC_INIT \
00125 asm( "movq %0, %%rsi " :: "m" (s)); \
00126 asm( "movq %0, %%rdi " :: "m" (d)); \
00127 asm( "movq %0, %%rcx " :: "m" (c)); \
00128 asm( "movq %0, %%rbx " :: "m" (b)); \
00129 asm( "xorq %r8, %r8 " );
00130
00131 #define MULADDC_CORE \
00132 asm( "movq (%rsi),%rax " ); \
00133 asm( "mulq %rbx " ); \
00134 asm( "addq $8, %rsi " ); \
00135 asm( "addq %rcx, %rax " ); \
00136 asm( "movq %r8, %rcx " ); \
00137 asm( "adcq $0, %rdx " ); \
00138 asm( "nop " ); \
00139 asm( "addq %rax, (%rdi) " ); \
00140 asm( "adcq %rdx, %rcx " ); \
00141 asm( "addq $8, %rdi " );
00142
00143 #define MULADDC_STOP \
00144 asm( "movq %%rcx, %0 " : "=m" (c)); \
00145 asm( "movq %%rdi, %0 " : "=m" (d)); \
00146 asm( "movq %%rsi, %0 " : "=m" (s) :: \
00147 "rax", "rcx", "rdx", "rbx", "rsi", "rdi", "r8" );
00148
00149 #endif
00150
00151 #if defined(__mc68020__) || defined(__mcpu32__)
00152
00153 #define MULADDC_INIT \
00154 asm( "movl %0, %%a2 " :: "m" (s)); \
00155 asm( "movl %0, %%a3 " :: "m" (d)); \
00156 asm( "movl %0, %%d3 " :: "m" (c)); \
00157 asm( "movl %0, %%d2 " :: "m" (b)); \
00158 asm( "moveq #0, %d0 " );
00159
00160 #define MULADDC_CORE \
00161 asm( "movel %a2@+, %d1 " ); \
00162 asm( "mulul %d2, %d4:%d1 " ); \
00163 asm( "addl %d3, %d1 " ); \
00164 asm( "addxl %d0, %d4 " ); \
00165 asm( "moveq #0, %d3 " ); \
00166 asm( "addl %d1, %a3@+ " ); \
00167 asm( "addxl %d4, %d3 " );
00168
00169 #define MULADDC_STOP \
00170 asm( "movl %%d3, %0 " : "=m" (c)); \
00171 asm( "movl %%a3, %0 " : "=m" (d)); \
00172 asm( "movl %%a2, %0 " : "=m" (s) :: \
00173 "d0", "d1", "d2", "d3", "d4", "a2", "a3" );
00174
00175 #define MULADDC_HUIT \
00176 asm( "movel %a2@+, %d1 " ); \
00177 asm( "mulul %d2, %d4:%d1 " ); \
00178 asm( "addxl %d3, %d1 " ); \
00179 asm( "addxl %d0, %d4 " ); \
00180 asm( "addl %d1, %a3@+ " ); \
00181 asm( "movel %a2@+, %d1 " ); \
00182 asm( "mulul %d2, %d3:%d1 " ); \
00183 asm( "addxl %d4, %d1 " ); \
00184 asm( "addxl %d0, %d3 " ); \
00185 asm( "addl %d1, %a3@+ " ); \
00186 asm( "movel %a2@+, %d1 " ); \
00187 asm( "mulul %d2, %d4:%d1 " ); \
00188 asm( "addxl %d3, %d1 " ); \
00189 asm( "addxl %d0, %d4 " ); \
00190 asm( "addl %d1, %a3@+ " ); \
00191 asm( "movel %a2@+, %d1 " ); \
00192 asm( "mulul %d2, %d3:%d1 " ); \
00193 asm( "addxl %d4, %d1 " ); \
00194 asm( "addxl %d0, %d3 " ); \
00195 asm( "addl %d1, %a3@+ " ); \
00196 asm( "movel %a2@+, %d1 " ); \
00197 asm( "mulul %d2, %d4:%d1 " ); \
00198 asm( "addxl %d3, %d1 " ); \
00199 asm( "addxl %d0, %d4 " ); \
00200 asm( "addl %d1, %a3@+ " ); \
00201 asm( "movel %a2@+, %d1 " ); \
00202 asm( "mulul %d2, %d3:%d1 " ); \
00203 asm( "addxl %d4, %d1 " ); \
00204 asm( "addxl %d0, %d3 " ); \
00205 asm( "addl %d1, %a3@+ " ); \
00206 asm( "movel %a2@+, %d1 " ); \
00207 asm( "mulul %d2, %d4:%d1 " ); \
00208 asm( "addxl %d3, %d1 " ); \
00209 asm( "addxl %d0, %d4 " ); \
00210 asm( "addl %d1, %a3@+ " ); \
00211 asm( "movel %a2@+, %d1 " ); \
00212 asm( "mulul %d2, %d3:%d1 " ); \
00213 asm( "addxl %d4, %d1 " ); \
00214 asm( "addxl %d0, %d3 " ); \
00215 asm( "addl %d1, %a3@+ " ); \
00216 asm( "addxl %d0, %d3 " );
00217
00218 #endif
00219
00220 #if defined(__powerpc__) || defined(__ppc__)
00221 #if defined(__powerpc64__) || defined(__ppc64__)
00222
00223 #define MULADDC_INIT \
00224 asm( "ld %%r3, %0 " :: "m" (s)); \
00225 asm( "ld %%r4, %0 " :: "m" (d)); \
00226 asm( "ld %%r5, %0 " :: "m" (c)); \
00227 asm( "ld %%r6, %0 " :: "m" (b)); \
00228 asm( "addi %r3, %r3, -8 " ); \
00229 asm( "addi %r4, %r4, -8 " ); \
00230 asm( "addic %r5, %r5, 0 " );
00231
00232 #define MULADDC_CORE \
00233 asm( "ldu %r7, 8(%r3) " ); \
00234 asm( "mulld %r8, %r7, %r6 " ); \
00235 asm( "mulhdu %r9, %r7, %r6 " ); \
00236 asm( "adde %r8, %r8, %r5 " ); \
00237 asm( "ld %r7, 8(%r4) " ); \
00238 asm( "addze %r5, %r9 " ); \
00239 asm( "addc %r8, %r8, %r7 " ); \
00240 asm( "stdu %r8, 8(%r4) " );
00241
00242 #define MULADDC_STOP \
00243 asm( "addze %r5, %r5 " ); \
00244 asm( "addi %r4, %r4, 8 " ); \
00245 asm( "addi %r3, %r3, 8 " ); \
00246 asm( "std %%r5, %0 " : "=m" (c)); \
00247 asm( "std %%r4, %0 " : "=m" (d)); \
00248 asm( "std %%r3, %0 " : "=m" (s) :: \
00249 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
00250
00251 #else
00252
00253 #define MULADDC_INIT \
00254 asm( "lwz %%r3, %0 " :: "m" (s)); \
00255 asm( "lwz %%r4, %0 " :: "m" (d)); \
00256 asm( "lwz %%r5, %0 " :: "m" (c)); \
00257 asm( "lwz %%r6, %0 " :: "m" (b)); \
00258 asm( "addi %r3, %r3, -4 " ); \
00259 asm( "addi %r4, %r4, -4 " ); \
00260 asm( "addic %r5, %r5, 0 " );
00261
00262 #define MULADDC_CORE \
00263 asm( "lwzu %r7, 4(%r3) " ); \
00264 asm( "mullw %r8, %r7, %r6 " ); \
00265 asm( "mulhwu %r9, %r7, %r6 " ); \
00266 asm( "adde %r8, %r8, %r5 " ); \
00267 asm( "lwz %r7, 4(%r4) " ); \
00268 asm( "addze %r5, %r9 " ); \
00269 asm( "addc %r8, %r8, %r7 " ); \
00270 asm( "stwu %r8, 4(%r4) " );
00271
00272 #define MULADDC_STOP \
00273 asm( "addze %r5, %r5 " ); \
00274 asm( "addi %r4, %r4, 4 " ); \
00275 asm( "addi %r3, %r3, 4 " ); \
00276 asm( "stw %%r5, %0 " : "=m" (c)); \
00277 asm( "stw %%r4, %0 " : "=m" (d)); \
00278 asm( "stw %%r3, %0 " : "=m" (s) :: \
00279 "r3", "r4", "r5", "r6", "r7", "r8", "r9" );
00280
00281 #endif
00282 #endif
00283
00284 #if defined(__sparc__)
00285
00286 #define MULADDC_INIT \
00287 asm( "ld %0, %%o0 " :: "m" (s)); \
00288 asm( "ld %0, %%o1 " :: "m" (d)); \
00289 asm( "ld %0, %%o2 " :: "m" (c)); \
00290 asm( "ld %0, %%o3 " :: "m" (b));
00291
00292 #define MULADDC_CORE \
00293 asm( "ld [%o0], %o4 " ); \
00294 asm( "inc 4, %o0 " ); \
00295 asm( "ld [%o1], %o5 " ); \
00296 asm( "umul %o3, %o4, %o4 " ); \
00297 asm( "addcc %o4, %o2, %o4 " ); \
00298 asm( "rd %y, %g1 " ); \
00299 asm( "addx %g1, 0, %g1 " ); \
00300 asm( "addcc %o4, %o5, %o4 " ); \
00301 asm( "st %o4, [%o1] " ); \
00302 asm( "addx %g1, 0, %o2 " ); \
00303 asm( "inc 4, %o1 " );
00304
00305 #define MULADDC_STOP \
00306 asm( "st %%o2, %0 " : "=m" (c)); \
00307 asm( "st %%o1, %0 " : "=m" (d)); \
00308 asm( "st %%o0, %0 " : "=m" (s) :: \
00309 "g1", "o0", "o1", "o2", "o3", "o4", "o5" );
00310
00311 #endif
00312
00313 #if defined(__microblaze__) || defined(microblaze)
00314
00315 #define MULADDC_INIT \
00316 asm( "lwi r3, %0 " :: "m" (s)); \
00317 asm( "lwi r4, %0 " :: "m" (d)); \
00318 asm( "lwi r5, %0 " :: "m" (c)); \
00319 asm( "lwi r6, %0 " :: "m" (b)); \
00320 asm( "andi r7, r6, 0xffff" ); \
00321 asm( "bsrli r6, r6, 16 " );
00322
00323 #define MULADDC_CORE \
00324 asm( "lhui r8, r3, 0 " ); \
00325 asm( "addi r3, r3, 2 " ); \
00326 asm( "lhui r9, r3, 0 " ); \
00327 asm( "addi r3, r3, 2 " ); \
00328 asm( "mul r10, r9, r6 " ); \
00329 asm( "mul r11, r8, r7 " ); \
00330 asm( "mul r12, r9, r7 " ); \
00331 asm( "mul r13, r8, r6 " ); \
00332 asm( "bsrli r8, r10, 16 " ); \
00333 asm( "bsrli r9, r11, 16 " ); \
00334 asm( "add r13, r13, r8 " ); \
00335 asm( "add r13, r13, r9 " ); \
00336 asm( "bslli r10, r10, 16 " ); \
00337 asm( "bslli r11, r11, 16 " ); \
00338 asm( "add r12, r12, r10 " ); \
00339 asm( "addc r13, r13, r0 " ); \
00340 asm( "add r12, r12, r11 " ); \
00341 asm( "addc r13, r13, r0 " ); \
00342 asm( "lwi r10, r4, 0 " ); \
00343 asm( "add r12, r12, r10 " ); \
00344 asm( "addc r13, r13, r0 " ); \
00345 asm( "add r12, r12, r5 " ); \
00346 asm( "addc r5, r13, r0 " ); \
00347 asm( "swi r12, r4, 0 " ); \
00348 asm( "addi r4, r4, 4 " );
00349
00350 #define MULADDC_STOP \
00351 asm( "swi r5, %0 " : "=m" (c)); \
00352 asm( "swi r4, %0 " : "=m" (d)); \
00353 asm( "swi r3, %0 " : "=m" (s) :: \
00354 "r3", "r4" , "r5" , "r6" , "r7" , "r8" , \
00355 "r9", "r10", "r11", "r12", "r13" );
00356
00357 #endif
00358
00359 #if defined(__tricore__)
00360
00361 #define MULADDC_INIT \
00362 asm( "ld.a %%a2, %0 " :: "m" (s)); \
00363 asm( "ld.a %%a3, %0 " :: "m" (d)); \
00364 asm( "ld.w %%d4, %0 " :: "m" (c)); \
00365 asm( "ld.w %%d1, %0 " :: "m" (b)); \
00366 asm( "xor %d5, %d5 " );
00367
00368 #define MULADDC_CORE \
00369 asm( "ld.w %d0, [%a2+] " ); \
00370 asm( "madd.u %e2, %e4, %d0, %d1 " ); \
00371 asm( "ld.w %d0, [%a3] " ); \
00372 asm( "addx %d2, %d2, %d0 " ); \
00373 asm( "addc %d3, %d3, 0 " ); \
00374 asm( "mov %d4, %d3 " ); \
00375 asm( "st.w [%a3+], %d2 " );
00376
00377 #define MULADDC_STOP \
00378 asm( "st.w %0, %%d4 " : "=m" (c)); \
00379 asm( "st.a %0, %%a3 " : "=m" (d)); \
00380 asm( "st.a %0, %%a2 " : "=m" (s) :: \
00381 "d0", "d1", "e2", "d4", "a2", "a3" );
00382
00383 #endif
00384
00385 #if defined(__arm__)
00386
00387 #define MULADDC_INIT \
00388 asm( "ldr r0, %0 " :: "m" (s)); \
00389 asm( "ldr r1, %0 " :: "m" (d)); \
00390 asm( "ldr r2, %0 " :: "m" (c)); \
00391 asm( "ldr r3, %0 " :: "m" (b));
00392
00393 #define MULADDC_CORE \
00394 asm( "ldr r4, [r0], #4 " ); \
00395 asm( "mov r5, #0 " ); \
00396 asm( "ldr r6, [r1] " ); \
00397 asm( "umlal r2, r5, r3, r4 " ); \
00398 asm( "adds r7, r6, r2 " ); \
00399 asm( "adc r2, r5, #0 " ); \
00400 asm( "str r7, [r1], #4 " );
00401
00402 #define MULADDC_STOP \
00403 asm( "str r2, %0 " : "=m" (c)); \
00404 asm( "str r1, %0 " : "=m" (d)); \
00405 asm( "str r0, %0 " : "=m" (s) :: \
00406 "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7" );
00407
00408 #endif
00409
00410 #if defined(__alpha__)
00411
00412 #define MULADDC_INIT \
00413 asm( "ldq $1, %0 " :: "m" (s)); \
00414 asm( "ldq $2, %0 " :: "m" (d)); \
00415 asm( "ldq $3, %0 " :: "m" (c)); \
00416 asm( "ldq $4, %0 " :: "m" (b));
00417
00418 #define MULADDC_CORE \
00419 asm( "ldq $6, 0($1) " ); \
00420 asm( "addq $1, 8, $1 " ); \
00421 asm( "mulq $6, $4, $7 " ); \
00422 asm( "umulh $6, $4, $6 " ); \
00423 asm( "addq $7, $3, $7 " ); \
00424 asm( "cmpult $7, $3, $3 " ); \
00425 asm( "ldq $5, 0($2) " ); \
00426 asm( "addq $7, $5, $7 " ); \
00427 asm( "cmpult $7, $5, $5 " ); \
00428 asm( "stq $7, 0($2) " ); \
00429 asm( "addq $2, 8, $2 " ); \
00430 asm( "addq $6, $3, $3 " ); \
00431 asm( "addq $5, $3, $3 " );
00432
00433 #define MULADDC_STOP \
00434 asm( "stq $3, %0 " : "=m" (c)); \
00435 asm( "stq $2, %0 " : "=m" (d)); \
00436 asm( "stq $1, %0 " : "=m" (s) :: \
00437 "$1", "$2", "$3", "$4", "$5", "$6", "$7" );
00438
00439 #endif
00440
00441 #if defined(__mips__)
00442
00443 #define MULADDC_INIT \
00444 asm( "lw $10, %0 " :: "m" (s)); \
00445 asm( "lw $11, %0 " :: "m" (d)); \
00446 asm( "lw $12, %0 " :: "m" (c)); \
00447 asm( "lw $13, %0 " :: "m" (b));
00448
00449 #define MULADDC_CORE \
00450 asm( "lw $14, 0($10) " ); \
00451 asm( "multu $13, $14 " ); \
00452 asm( "addi $10, $10, 4 " ); \
00453 asm( "mflo $14 " ); \
00454 asm( "mfhi $9 " ); \
00455 asm( "addu $14, $12, $14 " ); \
00456 asm( "lw $15, 0($11) " ); \
00457 asm( "sltu $12, $14, $12 " ); \
00458 asm( "addu $15, $14, $15 " ); \
00459 asm( "sltu $14, $15, $14 " ); \
00460 asm( "addu $12, $12, $9 " ); \
00461 asm( "sw $15, 0($11) " ); \
00462 asm( "addu $12, $12, $14 " ); \
00463 asm( "addi $11, $11, 4 " );
00464
00465 #define MULADDC_STOP \
00466 asm( "sw $12, %0 " : "=m" (c)); \
00467 asm( "sw $11, %0 " : "=m" (d)); \
00468 asm( "sw $10, %0 " : "=m" (s) :: \
00469 "$9", "$10", "$11", "$12", "$13", "$14", "$15" );
00470
00471 #endif
00472 #endif
00473
00474 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
00475
00476 #define MULADDC_INIT \
00477 __asm mov esi, s \
00478 __asm mov edi, d \
00479 __asm mov ecx, c \
00480 __asm mov ebx, b
00481
00482 #define MULADDC_CORE \
00483 __asm lodsd \
00484 __asm mul ebx \
00485 __asm add eax, ecx \
00486 __asm adc edx, 0 \
00487 __asm add eax, [edi] \
00488 __asm adc edx, 0 \
00489 __asm mov ecx, edx \
00490 __asm stosd
00491
00492 #define MULADDC_STOP \
00493 __asm mov c, ecx \
00494 __asm mov d, edi \
00495 __asm mov s, esi \
00496
00497 #if defined HAVE_SSE2
00498
00499 #define EMIT __asm _emit
00500
00501 #define MULADDC_HUIT \
00502 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
00503 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
00504 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
00505 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
00506 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
00507 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
00508 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
00509 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
00510 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
00511 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
00512 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
00513 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
00514 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
00515 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
00516 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
00517 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
00518 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
00519 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
00520 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
00521 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
00522 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
00523 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
00524 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00525 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
00526 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
00527 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
00528 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
00529 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
00530 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
00531 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00532 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
00533 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
00534 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
00535 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
00536 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
00537 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
00538 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00539 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
00540 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
00541 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
00542 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
00543 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00544 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
00545 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
00546 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
00547 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
00548 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00549 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
00550 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
00551 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
00552 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
00553 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00554 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
00555 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
00556 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00557 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
00558 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
00559 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
00560 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
00561 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
00562 EMIT 0x0F EMIT 0x7E EMIT 0xC9
00563
00564 #endif
00565 #endif
00566
00567 #if !defined(MULADDC_CORE)
00568 #if defined(HAVE_LONGLONG)
00569
00570 #define MULADDC_INIT \
00571 { \
00572 t_dbl r; \
00573 t_int r0, r1;
00574
00575 #define MULADDC_CORE \
00576 r = *(s++) * (t_dbl) b; \
00577 r0 = r; \
00578 r1 = r >> biL; \
00579 r0 += c; r1 += (r0 < c); \
00580 r0 += *d; r1 += (r0 < *d); \
00581 c = r1; *(d++) = r0;
00582
00583 #define MULADDC_STOP \
00584 }
00585
00586 #else
00587 #define MULADDC_INIT \
00588 { \
00589 t_int s0, s1, b0, b1; \
00590 t_int r0, r1, rx, ry; \
00591 b0 = ( b << biH ) >> biH; \
00592 b1 = ( b >> biH );
00593
00594 #define MULADDC_CORE \
00595 s0 = ( *s << biH ) >> biH; \
00596 s1 = ( *s >> biH ); s++; \
00597 rx = s0 * b1; r0 = s0 * b0; \
00598 ry = s1 * b0; r1 = s1 * b1; \
00599 r1 += ( rx >> biH ); \
00600 r1 += ( ry >> biH ); \
00601 rx <<= biH; ry <<= biH; \
00602 r0 += rx; r1 += (r0 < rx); \
00603 r0 += ry; r1 += (r0 < ry); \
00604 r0 += c; r1 += (r0 < c); \
00605 r0 += *d; r1 += (r0 < *d); \
00606 c = r1; *(d++) = r0;
00607
00608 #define MULADDC_STOP \
00609 }
00610
00611 #endif
00612 #endif
00613
00614 #endif
