bn/asm/x86_64-gcc.c

309 /* mul_add_c(a,b,c0,c1,c2)  -- c+=a*b for three word number c=(c2,c1,c0) */
310 /* mul_add_c2(a,b,c0,c1,c2) -- c+=2*a*b for three word number c=(c2,c1,c0) */
311 /* sqr_add_c(a,i,c0,c1,c2)  -- c+=a[i]^2 for three word number c=(c2,c1,c0) */
313  * sqr_add_c2(a,i,c0,c1,c2) -- c+=2*a[i]*a[j] for three word number
314  * c=(c2,c1,c0)
323 #  define mul_add_c(a,b,c0,c1,c2)       do {    \
328         c1 += hi; c2 += (c1<hi)?1:0;            \
331 #  define mul_add_c2(a,b,c0,c1,c2)      do {    \
336         c1 += tt; c2 += (c1<tt)?1:0;            \
338         c1 += hi; c2 += (c1<hi)?1:0;            \
341 #  define sqr_add_c(a,i,c0,c1,c2)       do {    \
346         c1 += hi; c2 += (c1<hi)?1:0;            \
349 #  define mul_add_c(a,b,c0,c1,c2) do {  \
356                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
361 #  define sqr_add_c(a,i,c0,c1,c2) do {  \
368                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
373 #  define mul_add_c2(a,b,c0,c1,c2) do { \
380                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
384                 : "+r"(c0),"+r"(c1),"+r"(c2)            \
390 # define sqr_add_c2(a,i,j,c0,c1,c2)      \
391         mul_add_c2((a)[i],(a)[j],c0,c1,c2)
395     BN_ULONG c1, c2, c3;
398     c2 = 0;
400     mul_add_c(a[0], b[0], c1, c2, c3);
403     mul_add_c(a[0], b[1], c2, c3, c1);
404     mul_add_c(a[1], b[0], c2, c3, c1);
405     r[1] = c2;
406     c2 = 0;
407     mul_add_c(a[2], b[0], c3, c1, c2);
408     mul_add_c(a[1], b[1], c3, c1, c2);
409     mul_add_c(a[0], b[2], c3, c1, c2);
412     mul_add_c(a[0], b[3], c1, c2, c3);
413     mul_add_c(a[1], b[2], c1, c2, c3);
414     mul_add_c(a[2], b[1], c1, c2, c3);
415     mul_add_c(a[3], b[0], c1, c2, c3);
418     mul_add_c(a[4], b[0], c2, c3, c1);
419     mul_add_c(a[3], b[1], c2, c3, c1);
420     mul_add_c(a[2], b[2], c2, c3, c1);
421     mul_add_c(a[1], b[3], c2, c3, c1);
422     mul_add_c(a[0], b[4], c2, c3, c1);
423     r[4] = c2;
424     c2 = 0;
425     mul_add_c(a[0], b[5], c3, c1, c2);
426     mul_add_c(a[1], b[4], c3, c1, c2);
427     mul_add_c(a[2], b[3], c3, c1, c2);
428     mul_add_c(a[3], b[2], c3, c1, c2);
429     mul_add_c(a[4], b[1], c3, c1, c2);
430     mul_add_c(a[5], b[0], c3, c1, c2);
433     mul_add_c(a[6], b[0], c1, c2, c3);
434     mul_add_c(a[5], b[1], c1, c2, c3);
435     mul_add_c(a[4], b[2], c1, c2, c3);
436     mul_add_c(a[3], b[3], c1, c2, c3);
437     mul_add_c(a[2], b[4], c1, c2, c3);
438     mul_add_c(a[1], b[5], c1, c2, c3);
439     mul_add_c(a[0], b[6], c1, c2, c3);
442     mul_add_c(a[0], b[7], c2, c3, c1);
443     mul_add_c(a[1], b[6], c2, c3, c1);
444     mul_add_c(a[2], b[5], c2, c3, c1);
445     mul_add_c(a[3], b[4], c2, c3, c1);
446     mul_add_c(a[4], b[3], c2, c3, c1);
447     mul_add_c(a[5], b[2], c2, c3, c1);
448     mul_add_c(a[6], b[1], c2, c3, c1);
449     mul_add_c(a[7], b[0], c2, c3, c1);
450     r[7] = c2;
451     c2 = 0;
452     mul_add_c(a[7], b[1], c3, c1, c2);
453     mul_add_c(a[6], b[2], c3, c1, c2);
454     mul_add_c(a[5], b[3], c3, c1, c2);
455     mul_add_c(a[4], b[4], c3, c1, c2);
456     mul_add_c(a[3], b[5], c3, c1, c2);
457     mul_add_c(a[2], b[6], c3, c1, c2);
458     mul_add_c(a[1], b[7], c3, c1, c2);
461     mul_add_c(a[2], b[7], c1, c2, c3);
462     mul_add_c(a[3], b[6], c1, c2, c3);
463     mul_add_c(a[4], b[5], c1, c2, c3);
464     mul_add_c(a[5], b[4], c1, c2, c3);
465     mul_add_c(a[6], b[3], c1, c2, c3);
466     mul_add_c(a[7], b[2], c1, c2, c3);
469     mul_add_c(a[7], b[3], c2, c3, c1);
470     mul_add_c(a[6], b[4], c2, c3, c1);
471     mul_add_c(a[5], b[5], c2, c3, c1);
472     mul_add_c(a[4], b[6], c2, c3, c1);
473     mul_add_c(a[3], b[7], c2, c3, c1);
474     r[10] = c2;
475     c2 = 0;
476     mul_add_c(a[4], b[7], c3, c1, c2);
477     mul_add_c(a[5], b[6], c3, c1, c2);
478     mul_add_c(a[6], b[5], c3, c1, c2);
479     mul_add_c(a[7], b[4], c3, c1, c2);
482     mul_add_c(a[7], b[5], c1, c2, c3);
483     mul_add_c(a[6], b[6], c1, c2, c3);
484     mul_add_c(a[5], b[7], c1, c2, c3);
487     mul_add_c(a[6], b[7], c2, c3, c1);
488     mul_add_c(a[7], b[6], c2, c3, c1);
489     r[13] = c2;
490     c2 = 0;
491     mul_add_c(a[7], b[7], c3, c1, c2);
498     BN_ULONG c1, c2, c3;
501     c2 = 0;
503     mul_add_c(a[0], b[0], c1, c2, c3);
506     mul_add_c(a[0], b[1], c2, c3, c1);
507     mul_add_c(a[1], b[0], c2, c3, c1);
508     r[1] = c2;
509     c2 = 0;
510     mul_add_c(a[2], b[0], c3, c1, c2);
511     mul_add_c(a[1], b[1], c3, c1, c2);
512     mul_add_c(a[0], b[2], c3, c1, c2);
515     mul_add_c(a[0], b[3], c1, c2, c3);
516     mul_add_c(a[1], b[2], c1, c2, c3);
517     mul_add_c(a[2], b[1], c1, c2, c3);
518     mul_add_c(a[3], b[0], c1, c2, c3);
521     mul_add_c(a[3], b[1], c2, c3, c1);
522     mul_add_c(a[2], b[2], c2, c3, c1);
523     mul_add_c(a[1], b[3], c2, c3, c1);
524     r[4] = c2;
525     c2 = 0;
526     mul_add_c(a[2], b[3], c3, c1, c2);
527     mul_add_c(a[3], b[2], c3, c1, c2);
530     mul_add_c(a[3], b[3], c1, c2, c3);
532     r[7] = c2;
537     BN_ULONG c1, c2, c3;
540     c2 = 0;
542     sqr_add_c(a, 0, c1, c2, c3);
545     sqr_add_c2(a, 1, 0, c2, c3, c1);
546     r[1] = c2;
547     c2 = 0;
548     sqr_add_c(a, 1, c3, c1, c2);
549     sqr_add_c2(a, 2, 0, c3, c1, c2);
552     sqr_add_c2(a, 3, 0, c1, c2, c3);
553     sqr_add_c2(a, 2, 1, c1, c2, c3);
556     sqr_add_c(a, 2, c2, c3, c1);
557     sqr_add_c2(a, 3, 1, c2, c3, c1);
558     sqr_add_c2(a, 4, 0, c2, c3, c1);
559     r[4] = c2;
560     c2 = 0;
561     sqr_add_c2(a, 5, 0, c3, c1, c2);
562     sqr_add_c2(a, 4, 1, c3, c1, c2);
563     sqr_add_c2(a, 3, 2, c3, c1, c2);
566     sqr_add_c(a, 3, c1, c2, c3);
567     sqr_add_c2(a, 4, 2, c1, c2, c3);
568     sqr_add_c2(a, 5, 1, c1, c2, c3);
569     sqr_add_c2(a, 6, 0, c1, c2, c3);
572     sqr_add_c2(a, 7, 0, c2, c3, c1);
573     sqr_add_c2(a, 6, 1, c2, c3, c1);
574     sqr_add_c2(a, 5, 2, c2, c3, c1);
575     sqr_add_c2(a, 4, 3, c2, c3, c1);
576     r[7] = c2;
577     c2 = 0;
578     sqr_add_c(a, 4, c3, c1, c2);
579     sqr_add_c2(a, 5, 3, c3, c1, c2);
580     sqr_add_c2(a, 6, 2, c3, c1, c2);
581     sqr_add_c2(a, 7, 1, c3, c1, c2);
584     sqr_add_c2(a, 7, 2, c1, c2, c3);
585     sqr_add_c2(a, 6, 3, c1, c2, c3);
586     sqr_add_c2(a, 5, 4, c1, c2, c3);
589     sqr_add_c(a, 5, c2, c3, c1);
590     sqr_add_c2(a, 6, 4, c2, c3, c1);
591     sqr_add_c2(a, 7, 3, c2, c3, c1);
592     r[10] = c2;
593     c2 = 0;
594     sqr_add_c2(a, 7, 4, c3, c1, c2);
595     sqr_add_c2(a, 6, 5, c3, c1, c2);
598     sqr_add_c(a, 6, c1, c2, c3);
599     sqr_add_c2(a, 7, 5, c1, c2, c3);
602     sqr_add_c2(a, 7, 6, c2, c3, c1);
603     r[13] = c2;
604     c2 = 0;
605     sqr_add_c(a, 7, c3, c1, c2);
612     BN_ULONG c1, c2, c3;
615     c2 = 0;
617     sqr_add_c(a, 0, c1, c2, c3);
620     sqr_add_c2(a, 1, 0, c2, c3, c1);
621     r[1] = c2;
622     c2 = 0;
623     sqr_add_c(a, 1, c3, c1, c2);
624     sqr_add_c2(a, 2, 0, c3, c1, c2);
627     sqr_add_c2(a, 3, 0, c1, c2, c3);
628     sqr_add_c2(a, 2, 1, c1, c2, c3);
631     sqr_add_c(a, 2, c2, c3, c1);
632     sqr_add_c2(a, 3, 1, c2, c3, c1);
633     r[4] = c2;
634     c2 = 0;
635     sqr_add_c2(a, 3, 2, c3, c1, c2);
638     sqr_add_c(a, 3, c1, c2, c3);
640     r[7] = c2;