2 * Copyright 2002-2018 The OpenSSL Project Authors. All Rights Reserved.
4 * Licensed under the Apache License 2.0 (the "License"). You may not use
5 * this file except in compliance with the License. You can obtain a copy
6 * in the file LICENSE in the source distribution or at
7 * https://www.openssl.org/source/license.html
10 #include "../bn_local.h"
11 #if !(defined(__GNUC__) && __GNUC__>=2)
12 # include "../bn_asm.c" /* kind of dirty hack for Sun Studio */
15 * x86_64 BIGNUM accelerator version 0.1, December 2002.
17 * Implemented by Andy Polyakov <appro@openssl.org> for the OpenSSL
20 * Rights for redistribution and usage in source and binary forms are
21 * granted according to the License. Warranty of any kind is disclaimed.
23 * Q. Version 0.1? It doesn't sound like Andy, he used to assign real
24 * versions, like 1.0...
25 * A. Well, that's because this code is basically a quick-n-dirty
26 * proof-of-concept hack. As you can see it's implemented with
27 * inline assembler, which means that you're bound to GCC and that
28 * there might be enough room for further improvement.
30 * Q. Why inline assembler?
31 * A. x86_64 features own ABI which I'm not familiar with. This is
32 * why I decided to let the compiler take care of subroutine
33 * prologue/epilogue as well as register allocation. For reference.
34 * Win64 implements different ABI for AMD64, different from Linux.
36 * Q. How much faster does it get?
37 * A. 'apps/openssl speed rsa dsa' output with no-asm:
39 * sign verify sign/s verify/s
40 * rsa 512 bits 0.0006s 0.0001s 1683.8 18456.2
41 * rsa 1024 bits 0.0028s 0.0002s 356.0 6407.0
42 * rsa 2048 bits 0.0172s 0.0005s 58.0 1957.8
43 * rsa 4096 bits 0.1155s 0.0018s 8.7 555.6
44 * sign verify sign/s verify/s
45 * dsa 512 bits 0.0005s 0.0006s 2100.8 1768.3
46 * dsa 1024 bits 0.0014s 0.0018s 692.3 559.2
47 * dsa 2048 bits 0.0049s 0.0061s 204.7 165.0
49 * 'apps/openssl speed rsa dsa' output with this module:
51 * sign verify sign/s verify/s
52 * rsa 512 bits 0.0004s 0.0000s 2767.1 33297.9
53 * rsa 1024 bits 0.0012s 0.0001s 867.4 14674.7
54 * rsa 2048 bits 0.0061s 0.0002s 164.0 5270.0
55 * rsa 4096 bits 0.0384s 0.0006s 26.1 1650.8
56 * sign verify sign/s verify/s
57 * dsa 512 bits 0.0002s 0.0003s 4442.2 3786.3
58 * dsa 1024 bits 0.0005s 0.0007s 1835.1 1497.4
59 * dsa 2048 bits 0.0016s 0.0020s 620.4 504.6
61 * For the reference. IA-32 assembler implementation performs
62 * very much like 64-bit code compiled with no-asm on the same
70 * "m"(a), "+m"(r) is the way to favor DirectPath ยต-code;
71 * "g"(0) let the compiler to decide where does it
72 * want to keep the value of zero;
74 # define mul_add(r,a,word,carry) do { \
75 register BN_ULONG high,low; \
77 : "=a"(low),"=d"(high) \
80 asm ("addq %2,%0; adcq %3,%1" \
81 : "+r"(carry),"+d"(high)\
84 asm ("addq %2,%0; adcq %3,%1" \
85 : "+m"(r),"+d"(high) \
91 # define mul(r,a,word,carry) do { \
92 register BN_ULONG high,low; \
94 : "=a"(low),"=d"(high) \
97 asm ("addq %2,%0; adcq %3,%1" \
98 : "+r"(carry),"+d"(high)\
101 (r)=carry, carry=high; \
104 # define sqr(r0,r1,a) \
106 : "=a"(r0),"=d"(r1) \
110 BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num,
119 mul_add(rp[0], ap[0], w, c1);
120 mul_add(rp[1], ap[1], w, c1);
121 mul_add(rp[2], ap[2], w, c1);
122 mul_add(rp[3], ap[3], w, c1);
128 mul_add(rp[0], ap[0], w, c1);
131 mul_add(rp[1], ap[1], w, c1);
134 mul_add(rp[2], ap[2], w, c1);
141 BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w)
149 mul(rp[0], ap[0], w, c1);
150 mul(rp[1], ap[1], w, c1);
151 mul(rp[2], ap[2], w, c1);
152 mul(rp[3], ap[3], w, c1);
158 mul(rp[0], ap[0], w, c1);
161 mul(rp[1], ap[1], w, c1);
164 mul(rp[2], ap[2], w, c1);
169 void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n)
175 sqr(r[0], r[1], a[0]);
176 sqr(r[2], r[3], a[1]);
177 sqr(r[4], r[5], a[2]);
178 sqr(r[6], r[7], a[3]);
184 sqr(r[0], r[1], a[0]);
187 sqr(r[2], r[3], a[1]);
190 sqr(r[4], r[5], a[2]);
194 BN_ULONG bn_div_words(BN_ULONG h, BN_ULONG l, BN_ULONG d)
198 asm("divq %4":"=a"(ret), "=d"(waste)
199 : "a"(l), "d"(h), "r"(d)
205 BN_ULONG bn_add_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
214 asm volatile (" subq %0,%0 \n" /* clear carry */
217 "1: movq (%4,%2,8),%0 \n"
218 " adcq (%5,%2,8),%0 \n"
219 " movq %0,(%3,%2,8) \n"
224 :"=&r" (ret), "+c"(n), "+r"(i)
225 :"r"(rp), "r"(ap), "r"(bp)
232 BN_ULONG bn_sub_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
241 asm volatile (" subq %0,%0 \n" /* clear borrow */
244 "1: movq (%4,%2,8),%0 \n"
245 " sbbq (%5,%2,8),%0 \n"
246 " movq %0,(%3,%2,8) \n"
251 :"=&r" (ret), "+c"(n), "+r"(i)
252 :"r"(rp), "r"(ap), "r"(bp)
258 /* Simics 1.4<7 has buggy sbbq:-( */
259 # define BN_MASK2 0xffffffffffffffffL
260 BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
271 r[0] = (t1 - t2 - c) & BN_MASK2;
279 r[1] = (t1 - t2 - c) & BN_MASK2;
287 r[2] = (t1 - t2 - c) & BN_MASK2;
295 r[3] = (t1 - t2 - c) & BN_MASK2;
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
318 * Keep in mind that carrying into high part of multiplication result
319 * can not overflow, because it cannot be all-ones.
322 /* original macros are kept for reference purposes */
323 # define mul_add_c(a,b,c0,c1,c2) do { \
324 BN_ULONG ta = (a), tb = (b); \
326 BN_UMULT_LOHI(lo,hi,ta,tb); \
327 c0 += lo; hi += (c0<lo)?1:0; \
328 c1 += hi; c2 += (c1<hi)?1:0; \
331 # define mul_add_c2(a,b,c0,c1,c2) do { \
332 BN_ULONG ta = (a), tb = (b); \
333 BN_ULONG lo, hi, tt; \
334 BN_UMULT_LOHI(lo,hi,ta,tb); \
335 c0 += lo; tt = hi+((c0<lo)?1:0); \
336 c1 += tt; c2 += (c1<tt)?1:0; \
337 c0 += lo; hi += (c0<lo)?1:0; \
338 c1 += hi; c2 += (c1<hi)?1:0; \
341 # define sqr_add_c(a,i,c0,c1,c2) do { \
342 BN_ULONG ta = (a)[i]; \
344 BN_UMULT_LOHI(lo,hi,ta,ta); \
345 c0 += lo; hi += (c0<lo)?1:0; \
346 c1 += hi; c2 += (c1<hi)?1:0; \
349 # define mul_add_c(a,b,c0,c1,c2) do { \
352 : "=a"(t1),"=d"(t2) \
355 asm ("addq %3,%0; adcq %4,%1; adcq %5,%2" \
356 : "+r"(c0),"+r"(c1),"+r"(c2) \
357 : "r"(t1),"r"(t2),"g"(0) \
361 # define sqr_add_c(a,i,c0,c1,c2) do { \
364 : "=a"(t1),"=d"(t2) \
367 asm ("addq %3,%0; adcq %4,%1; adcq %5,%2" \
368 : "+r"(c0),"+r"(c1),"+r"(c2) \
369 : "r"(t1),"r"(t2),"g"(0) \
373 # define mul_add_c2(a,b,c0,c1,c2) do { \
376 : "=a"(t1),"=d"(t2) \
379 asm ("addq %3,%0; adcq %4,%1; adcq %5,%2" \
380 : "+r"(c0),"+r"(c1),"+r"(c2) \
381 : "r"(t1),"r"(t2),"g"(0) \
383 asm ("addq %3,%0; adcq %4,%1; adcq %5,%2" \
384 : "+r"(c0),"+r"(c1),"+r"(c2) \
385 : "r"(t1),"r"(t2),"g"(0) \
390 # define sqr_add_c2(a,i,j,c0,c1,c2) \
391 mul_add_c2((a)[i],(a)[j],c0,c1,c2)
393 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
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);
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);
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);
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);
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);
491 mul_add_c(a[7], b[7], c3, c1, c2);
496 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
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);
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);
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);
535 void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a)
542 sqr_add_c(a, 0, c1, c2, c3);
545 sqr_add_c2(a, 1, 0, c2, c3, c1);
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);
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);
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);
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);
605 sqr_add_c(a, 7, c3, c1, c2);
610 void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a)
617 sqr_add_c(a, 0, c1, c2, c3);
620 sqr_add_c2(a, 1, 0, c2, c3, c1);
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);
635 sqr_add_c2(a, 3, 2, c3, c1, c2);
638 sqr_add_c(a, 3, c1, c2, c3);