1 /* x86_64 BIGNUM accelerator version 0.1, December 2002.
2  *
3  * Implemented by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
4  * project.
5  *
6  * Rights for redistribution and usage in source and binary forms are
7  * granted according to the OpenSSL license. Warranty of any kind is
8  * disclaimed.
9  *
10  * Q. Version 0.1? It doesn't sound like Andy, he used to assign real
11  *    versions, like 1.0...
12  * A. Well, that's because this code is basically a quick-n-dirty
13  *    proof-of-concept hack. As you can see it's implemented with
14  *    inline assembler, which means that you're bound to GCC and that
15  *    there might be enough room for further improvement.
16  *
17  * Q. Why inline assembler?
18  * A. x86_64 features own ABI which I'm not familiar with. This is
19  *    why I decided to let the compiler take care of subroutine
20  *    prologue/epilogue as well as register allocation. For reference.
21  *    Win64 implements different ABI for AMD64, different from Linux.
22  *
23  * Q. How much faster does it get?
24  * A. 'apps/openssl speed rsa dsa' output with no-asm:
25  *
26  *	                  sign    verify    sign/s verify/s
27  *	rsa  512 bits   0.0006s   0.0001s   1683.8  18456.2
28  *	rsa 1024 bits   0.0028s   0.0002s    356.0   6407.0
29  *	rsa 2048 bits   0.0172s   0.0005s     58.0   1957.8
30  *	rsa 4096 bits   0.1155s   0.0018s      8.7    555.6
31  *	                  sign    verify    sign/s verify/s
32  *	dsa  512 bits   0.0005s   0.0006s   2100.8   1768.3
33  *	dsa 1024 bits   0.0014s   0.0018s    692.3    559.2
34  *	dsa 2048 bits   0.0049s   0.0061s    204.7    165.0
35  *
36  *    'apps/openssl speed rsa dsa' output with this module:
37  *
38  *	                  sign    verify    sign/s verify/s
39  *	rsa  512 bits   0.0004s   0.0000s   2767.1  33297.9
40  *	rsa 1024 bits   0.0012s   0.0001s    867.4  14674.7
41  *	rsa 2048 bits   0.0061s   0.0002s    164.0   5270.0
42  *	rsa 4096 bits   0.0384s   0.0006s     26.1   1650.8
43  *	                  sign    verify    sign/s verify/s
44  *	dsa  512 bits   0.0002s   0.0003s   4442.2   3786.3
45  *	dsa 1024 bits   0.0005s   0.0007s   1835.1   1497.4
46  *	dsa 2048 bits   0.0016s   0.0020s    620.4    504.6
47  *
48  *    For the reference. IA-32 assembler implementation performs
49  *    very much like 64-bit code compiled with no-asm on the same
50  *    machine.
51  */
52 
53 #include <openssl/bn.h>
54 
55 /* TODO(davidben): Get this file working on Windows x64. */
56 #if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64) && defined(__GNUC__)
57 
58 #include "../internal.h"
59 
60 
61 #undef mul
62 #undef mul_add
63 
64 #define asm __asm__
65 
66 /*
67  * "m"(a), "+m"(r)	is the way to favor DirectPath µ-code;
68  * "g"(0)		let the compiler to decide where does it
69  *			want to keep the value of zero;
70  */
71 #define mul_add(r, a, word, carry)                                     \
72   do {                                                                 \
73     register BN_ULONG high, low;                                       \
74     asm("mulq %3" : "=a"(low), "=d"(high) : "a"(word), "m"(a) : "cc"); \
75     asm("addq %2,%0; adcq %3,%1"                                       \
76         : "+r"(carry), "+d"(high)                                      \
77         : "a"(low), "g"(0)                                             \
78         : "cc");                                                       \
79     asm("addq %2,%0; adcq %3,%1"                                       \
80         : "+m"(r), "+d"(high)                                          \
81         : "r"(carry), "g"(0)                                           \
82         : "cc");                                                       \
83     (carry) = high;                                                    \
84   } while (0)
85 
86 #define mul(r, a, word, carry)                                         \
87   do {                                                                 \
88     register BN_ULONG high, low;                                       \
89     asm("mulq %3" : "=a"(low), "=d"(high) : "a"(word), "g"(a) : "cc"); \
90     asm("addq %2,%0; adcq %3,%1"                                       \
91         : "+r"(carry), "+d"(high)                                      \
92         : "a"(low), "g"(0)                                             \
93         : "cc");                                                       \
94     (r) = (carry);                                                     \
95     (carry) = high;                                                    \
96   } while (0)
97 #undef sqr
98 #define sqr(r0, r1, a) asm("mulq %2" : "=a"(r0), "=d"(r1) : "a"(a) : "cc");
99 
bn_mul_add_words(BN_ULONG * rp,const BN_ULONG * ap,int num,BN_ULONG w)100 BN_ULONG bn_mul_add_words(BN_ULONG *rp, const BN_ULONG *ap, int num,
101                           BN_ULONG w) {
102   BN_ULONG c1 = 0;
103 
104   if (num <= 0) {
105     return (c1);
106   }
107 
108   while (num & ~3) {
109     mul_add(rp[0], ap[0], w, c1);
110     mul_add(rp[1], ap[1], w, c1);
111     mul_add(rp[2], ap[2], w, c1);
112     mul_add(rp[3], ap[3], w, c1);
113     ap += 4;
114     rp += 4;
115     num -= 4;
116   }
117   if (num) {
118     mul_add(rp[0], ap[0], w, c1);
119     if (--num == 0) {
120       return c1;
121     }
122     mul_add(rp[1], ap[1], w, c1);
123     if (--num == 0) {
124       return c1;
125     }
126     mul_add(rp[2], ap[2], w, c1);
127     return c1;
128   }
129 
130   return c1;
131 }
132 
bn_mul_words(BN_ULONG * rp,const BN_ULONG * ap,int num,BN_ULONG w)133 BN_ULONG bn_mul_words(BN_ULONG *rp, const BN_ULONG *ap, int num, BN_ULONG w) {
134   BN_ULONG c1 = 0;
135 
136   if (num <= 0) {
137     return c1;
138   }
139 
140   while (num & ~3) {
141     mul(rp[0], ap[0], w, c1);
142     mul(rp[1], ap[1], w, c1);
143     mul(rp[2], ap[2], w, c1);
144     mul(rp[3], ap[3], w, c1);
145     ap += 4;
146     rp += 4;
147     num -= 4;
148   }
149   if (num) {
150     mul(rp[0], ap[0], w, c1);
151     if (--num == 0) {
152       return c1;
153     }
154     mul(rp[1], ap[1], w, c1);
155     if (--num == 0) {
156       return c1;
157     }
158     mul(rp[2], ap[2], w, c1);
159   }
160   return c1;
161 }
162 
bn_sqr_words(BN_ULONG * r,const BN_ULONG * a,int n)163 void bn_sqr_words(BN_ULONG *r, const BN_ULONG *a, int n) {
164   if (n <= 0) {
165     return;
166   }
167 
168   while (n & ~3) {
169     sqr(r[0], r[1], a[0]);
170     sqr(r[2], r[3], a[1]);
171     sqr(r[4], r[5], a[2]);
172     sqr(r[6], r[7], a[3]);
173     a += 4;
174     r += 8;
175     n -= 4;
176   }
177   if (n) {
178     sqr(r[0], r[1], a[0]);
179     if (--n == 0) {
180       return;
181     }
182     sqr(r[2], r[3], a[1]);
183     if (--n == 0) {
184       return;
185     }
186     sqr(r[4], r[5], a[2]);
187   }
188 }
189 
bn_add_words(BN_ULONG * rp,const BN_ULONG * ap,const BN_ULONG * bp,int n)190 BN_ULONG bn_add_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
191                       int n) {
192   BN_ULONG ret;
193   size_t i = 0;
194 
195   if (n <= 0) {
196     return 0;
197   }
198 
199   asm volatile (
200       "	subq	%0,%0		\n" /* clear carry */
201       "	jmp	1f		\n"
202       ".p2align 4			\n"
203       "1:	movq	(%4,%2,8),%0	\n"
204       "	adcq	(%5,%2,8),%0	\n"
205       "	movq	%0,(%3,%2,8)	\n"
206       "	lea	1(%2),%2	\n"
207       "	loop	1b		\n"
208       "	sbbq	%0,%0		\n"
209       : "=&r"(ret), "+c"(n), "+r"(i)
210       : "r"(rp), "r"(ap), "r"(bp)
211       : "cc", "memory");
212 
213   return ret & 1;
214 }
215 
bn_sub_words(BN_ULONG * rp,const BN_ULONG * ap,const BN_ULONG * bp,int n)216 BN_ULONG bn_sub_words(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp,
217                       int n) {
218   BN_ULONG ret;
219   size_t i = 0;
220 
221   if (n <= 0) {
222     return 0;
223   }
224 
225   asm volatile (
226       "	subq	%0,%0		\n" /* clear borrow */
227       "	jmp	1f		\n"
228       ".p2align 4			\n"
229       "1:	movq	(%4,%2,8),%0	\n"
230       "	sbbq	(%5,%2,8),%0	\n"
231       "	movq	%0,(%3,%2,8)	\n"
232       "	lea	1(%2),%2	\n"
233       "	loop	1b		\n"
234       "	sbbq	%0,%0		\n"
235       : "=&r"(ret), "+c"(n), "+r"(i)
236       : "r"(rp), "r"(ap), "r"(bp)
237       : "cc", "memory");
238 
239   return ret & 1;
240 }
241 
242 /* mul_add_c(a,b,c0,c1,c2)  -- c+=a*b for three word number c=(c2,c1,c0) */
243 /* mul_add_c2(a,b,c0,c1,c2) -- c+=2*a*b for three word number c=(c2,c1,c0) */
244 /* sqr_add_c(a,i,c0,c1,c2)  -- c+=a[i]^2 for three word number c=(c2,c1,c0) */
245 /* sqr_add_c2(a,i,c0,c1,c2) -- c+=2*a[i]*a[j] for three word number c=(c2,c1,c0)
246  */
247 
248 /* Keep in mind that carrying into high part of multiplication result can not
249  * overflow, because it cannot be all-ones. */
250 #define mul_add_c(a, b, c0, c1, c2)          \
251   do {                                       \
252     BN_ULONG t1, t2;                \
253     asm("mulq %3" : "=a"(t1), "=d"(t2) : "a"(a), "m"(b) : "cc"); \
254     asm("addq %3,%0; adcq %4,%1; adcq %5,%2" \
255         : "+r"(c0), "+r"(c1), "+r"(c2)       \
256         : "r"(t1), "r"(t2), "g"(0)           \
257         : "cc");                             \
258   } while (0)
259 
260 #define sqr_add_c(a, i, c0, c1, c2)                           \
261   do {                                                        \
262     BN_ULONG t1, t2;                                          \
263     asm("mulq %2" : "=a"(t1), "=d"(t2) : "a"((a)[i]) : "cc"); \
264     asm("addq %3,%0; adcq %4,%1; adcq %5,%2"                  \
265         : "+r"(c0), "+r"(c1), "+r"(c2)                        \
266         : "r"(t1), "r"(t2), "g"(0)                            \
267         : "cc");                                              \
268   } while (0)
269 
270 #define mul_add_c2(a, b, c0, c1, c2)         \
271   do {                                       \
272     BN_ULONG t1, t2;                                                    \
273     asm("mulq %3" : "=a"(t1), "=d"(t2) : "a"(a), "m"(b) : "cc");        \
274     asm("addq %3,%0; adcq %4,%1; adcq %5,%2" \
275         : "+r"(c0), "+r"(c1), "+r"(c2)       \
276         : "r"(t1), "r"(t2), "g"(0)           \
277         : "cc");                             \
278     asm("addq %3,%0; adcq %4,%1; adcq %5,%2" \
279         : "+r"(c0), "+r"(c1), "+r"(c2)       \
280         : "r"(t1), "r"(t2), "g"(0)           \
281         : "cc");                             \
282   } while (0)
283 
284 #define sqr_add_c2(a, i, j, c0, c1, c2) mul_add_c2((a)[i], (a)[j], c0, c1, c2)
285 
bn_mul_comba8(BN_ULONG * r,BN_ULONG * a,BN_ULONG * b)286 void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) {
287   BN_ULONG c1, c2, c3;
288 
289   c1 = 0;
290   c2 = 0;
291   c3 = 0;
292   mul_add_c(a[0], b[0], c1, c2, c3);
293   r[0] = c1;
294   c1 = 0;
295   mul_add_c(a[0], b[1], c2, c3, c1);
296   mul_add_c(a[1], b[0], c2, c3, c1);
297   r[1] = c2;
298   c2 = 0;
299   mul_add_c(a[2], b[0], c3, c1, c2);
300   mul_add_c(a[1], b[1], c3, c1, c2);
301   mul_add_c(a[0], b[2], c3, c1, c2);
302   r[2] = c3;
303   c3 = 0;
304   mul_add_c(a[0], b[3], c1, c2, c3);
305   mul_add_c(a[1], b[2], c1, c2, c3);
306   mul_add_c(a[2], b[1], c1, c2, c3);
307   mul_add_c(a[3], b[0], c1, c2, c3);
308   r[3] = c1;
309   c1 = 0;
310   mul_add_c(a[4], b[0], c2, c3, c1);
311   mul_add_c(a[3], b[1], c2, c3, c1);
312   mul_add_c(a[2], b[2], c2, c3, c1);
313   mul_add_c(a[1], b[3], c2, c3, c1);
314   mul_add_c(a[0], b[4], c2, c3, c1);
315   r[4] = c2;
316   c2 = 0;
317   mul_add_c(a[0], b[5], c3, c1, c2);
318   mul_add_c(a[1], b[4], c3, c1, c2);
319   mul_add_c(a[2], b[3], c3, c1, c2);
320   mul_add_c(a[3], b[2], c3, c1, c2);
321   mul_add_c(a[4], b[1], c3, c1, c2);
322   mul_add_c(a[5], b[0], c3, c1, c2);
323   r[5] = c3;
324   c3 = 0;
325   mul_add_c(a[6], b[0], c1, c2, c3);
326   mul_add_c(a[5], b[1], c1, c2, c3);
327   mul_add_c(a[4], b[2], c1, c2, c3);
328   mul_add_c(a[3], b[3], c1, c2, c3);
329   mul_add_c(a[2], b[4], c1, c2, c3);
330   mul_add_c(a[1], b[5], c1, c2, c3);
331   mul_add_c(a[0], b[6], c1, c2, c3);
332   r[6] = c1;
333   c1 = 0;
334   mul_add_c(a[0], b[7], c2, c3, c1);
335   mul_add_c(a[1], b[6], c2, c3, c1);
336   mul_add_c(a[2], b[5], c2, c3, c1);
337   mul_add_c(a[3], b[4], c2, c3, c1);
338   mul_add_c(a[4], b[3], c2, c3, c1);
339   mul_add_c(a[5], b[2], c2, c3, c1);
340   mul_add_c(a[6], b[1], c2, c3, c1);
341   mul_add_c(a[7], b[0], c2, c3, c1);
342   r[7] = c2;
343   c2 = 0;
344   mul_add_c(a[7], b[1], c3, c1, c2);
345   mul_add_c(a[6], b[2], c3, c1, c2);
346   mul_add_c(a[5], b[3], c3, c1, c2);
347   mul_add_c(a[4], b[4], c3, c1, c2);
348   mul_add_c(a[3], b[5], c3, c1, c2);
349   mul_add_c(a[2], b[6], c3, c1, c2);
350   mul_add_c(a[1], b[7], c3, c1, c2);
351   r[8] = c3;
352   c3 = 0;
353   mul_add_c(a[2], b[7], c1, c2, c3);
354   mul_add_c(a[3], b[6], c1, c2, c3);
355   mul_add_c(a[4], b[5], c1, c2, c3);
356   mul_add_c(a[5], b[4], c1, c2, c3);
357   mul_add_c(a[6], b[3], c1, c2, c3);
358   mul_add_c(a[7], b[2], c1, c2, c3);
359   r[9] = c1;
360   c1 = 0;
361   mul_add_c(a[7], b[3], c2, c3, c1);
362   mul_add_c(a[6], b[4], c2, c3, c1);
363   mul_add_c(a[5], b[5], c2, c3, c1);
364   mul_add_c(a[4], b[6], c2, c3, c1);
365   mul_add_c(a[3], b[7], c2, c3, c1);
366   r[10] = c2;
367   c2 = 0;
368   mul_add_c(a[4], b[7], c3, c1, c2);
369   mul_add_c(a[5], b[6], c3, c1, c2);
370   mul_add_c(a[6], b[5], c3, c1, c2);
371   mul_add_c(a[7], b[4], c3, c1, c2);
372   r[11] = c3;
373   c3 = 0;
374   mul_add_c(a[7], b[5], c1, c2, c3);
375   mul_add_c(a[6], b[6], c1, c2, c3);
376   mul_add_c(a[5], b[7], c1, c2, c3);
377   r[12] = c1;
378   c1 = 0;
379   mul_add_c(a[6], b[7], c2, c3, c1);
380   mul_add_c(a[7], b[6], c2, c3, c1);
381   r[13] = c2;
382   c2 = 0;
383   mul_add_c(a[7], b[7], c3, c1, c2);
384   r[14] = c3;
385   r[15] = c1;
386 }
387 
bn_mul_comba4(BN_ULONG * r,BN_ULONG * a,BN_ULONG * b)388 void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b) {
389   BN_ULONG c1, c2, c3;
390 
391   c1 = 0;
392   c2 = 0;
393   c3 = 0;
394   mul_add_c(a[0], b[0], c1, c2, c3);
395   r[0] = c1;
396   c1 = 0;
397   mul_add_c(a[0], b[1], c2, c3, c1);
398   mul_add_c(a[1], b[0], c2, c3, c1);
399   r[1] = c2;
400   c2 = 0;
401   mul_add_c(a[2], b[0], c3, c1, c2);
402   mul_add_c(a[1], b[1], c3, c1, c2);
403   mul_add_c(a[0], b[2], c3, c1, c2);
404   r[2] = c3;
405   c3 = 0;
406   mul_add_c(a[0], b[3], c1, c2, c3);
407   mul_add_c(a[1], b[2], c1, c2, c3);
408   mul_add_c(a[2], b[1], c1, c2, c3);
409   mul_add_c(a[3], b[0], c1, c2, c3);
410   r[3] = c1;
411   c1 = 0;
412   mul_add_c(a[3], b[1], c2, c3, c1);
413   mul_add_c(a[2], b[2], c2, c3, c1);
414   mul_add_c(a[1], b[3], c2, c3, c1);
415   r[4] = c2;
416   c2 = 0;
417   mul_add_c(a[2], b[3], c3, c1, c2);
418   mul_add_c(a[3], b[2], c3, c1, c2);
419   r[5] = c3;
420   c3 = 0;
421   mul_add_c(a[3], b[3], c1, c2, c3);
422   r[6] = c1;
423   r[7] = c2;
424 }
425 
bn_sqr_comba8(BN_ULONG * r,const BN_ULONG * a)426 void bn_sqr_comba8(BN_ULONG *r, const BN_ULONG *a) {
427   BN_ULONG c1, c2, c3;
428 
429   c1 = 0;
430   c2 = 0;
431   c3 = 0;
432   sqr_add_c(a, 0, c1, c2, c3);
433   r[0] = c1;
434   c1 = 0;
435   sqr_add_c2(a, 1, 0, c2, c3, c1);
436   r[1] = c2;
437   c2 = 0;
438   sqr_add_c(a, 1, c3, c1, c2);
439   sqr_add_c2(a, 2, 0, c3, c1, c2);
440   r[2] = c3;
441   c3 = 0;
442   sqr_add_c2(a, 3, 0, c1, c2, c3);
443   sqr_add_c2(a, 2, 1, c1, c2, c3);
444   r[3] = c1;
445   c1 = 0;
446   sqr_add_c(a, 2, c2, c3, c1);
447   sqr_add_c2(a, 3, 1, c2, c3, c1);
448   sqr_add_c2(a, 4, 0, c2, c3, c1);
449   r[4] = c2;
450   c2 = 0;
451   sqr_add_c2(a, 5, 0, c3, c1, c2);
452   sqr_add_c2(a, 4, 1, c3, c1, c2);
453   sqr_add_c2(a, 3, 2, c3, c1, c2);
454   r[5] = c3;
455   c3 = 0;
456   sqr_add_c(a, 3, c1, c2, c3);
457   sqr_add_c2(a, 4, 2, c1, c2, c3);
458   sqr_add_c2(a, 5, 1, c1, c2, c3);
459   sqr_add_c2(a, 6, 0, c1, c2, c3);
460   r[6] = c1;
461   c1 = 0;
462   sqr_add_c2(a, 7, 0, c2, c3, c1);
463   sqr_add_c2(a, 6, 1, c2, c3, c1);
464   sqr_add_c2(a, 5, 2, c2, c3, c1);
465   sqr_add_c2(a, 4, 3, c2, c3, c1);
466   r[7] = c2;
467   c2 = 0;
468   sqr_add_c(a, 4, c3, c1, c2);
469   sqr_add_c2(a, 5, 3, c3, c1, c2);
470   sqr_add_c2(a, 6, 2, c3, c1, c2);
471   sqr_add_c2(a, 7, 1, c3, c1, c2);
472   r[8] = c3;
473   c3 = 0;
474   sqr_add_c2(a, 7, 2, c1, c2, c3);
475   sqr_add_c2(a, 6, 3, c1, c2, c3);
476   sqr_add_c2(a, 5, 4, c1, c2, c3);
477   r[9] = c1;
478   c1 = 0;
479   sqr_add_c(a, 5, c2, c3, c1);
480   sqr_add_c2(a, 6, 4, c2, c3, c1);
481   sqr_add_c2(a, 7, 3, c2, c3, c1);
482   r[10] = c2;
483   c2 = 0;
484   sqr_add_c2(a, 7, 4, c3, c1, c2);
485   sqr_add_c2(a, 6, 5, c3, c1, c2);
486   r[11] = c3;
487   c3 = 0;
488   sqr_add_c(a, 6, c1, c2, c3);
489   sqr_add_c2(a, 7, 5, c1, c2, c3);
490   r[12] = c1;
491   c1 = 0;
492   sqr_add_c2(a, 7, 6, c2, c3, c1);
493   r[13] = c2;
494   c2 = 0;
495   sqr_add_c(a, 7, c3, c1, c2);
496   r[14] = c3;
497   r[15] = c1;
498 }
499 
bn_sqr_comba4(BN_ULONG * r,const BN_ULONG * a)500 void bn_sqr_comba4(BN_ULONG *r, const BN_ULONG *a) {
501   BN_ULONG c1, c2, c3;
502 
503   c1 = 0;
504   c2 = 0;
505   c3 = 0;
506   sqr_add_c(a, 0, c1, c2, c3);
507   r[0] = c1;
508   c1 = 0;
509   sqr_add_c2(a, 1, 0, c2, c3, c1);
510   r[1] = c2;
511   c2 = 0;
512   sqr_add_c(a, 1, c3, c1, c2);
513   sqr_add_c2(a, 2, 0, c3, c1, c2);
514   r[2] = c3;
515   c3 = 0;
516   sqr_add_c2(a, 3, 0, c1, c2, c3);
517   sqr_add_c2(a, 2, 1, c1, c2, c3);
518   r[3] = c1;
519   c1 = 0;
520   sqr_add_c(a, 2, c2, c3, c1);
521   sqr_add_c2(a, 3, 1, c2, c3, c1);
522   r[4] = c2;
523   c2 = 0;
524   sqr_add_c2(a, 3, 2, c3, c1, c2);
525   r[5] = c3;
526   c3 = 0;
527   sqr_add_c(a, 3, c1, c2, c3);
528   r[6] = c1;
529   r[7] = c2;
530 }
531 
532 #endif  /* !NO_ASM && X86_64 && __GNUC__ */
533