xref: /freebsd/contrib/bearssl/src/int/i31_moddiv.c (revision a90b9d0159070121c221b966469c3e36d912bf82)
1 /*
2  * Copyright (c) 2018 Thomas Pornin <pornin@bolet.org>
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining
5  * a copy of this software and associated documentation files (the
6  * "Software"), to deal in the Software without restriction, including
7  * without limitation the rights to use, copy, modify, merge, publish,
8  * distribute, sublicense, and/or sell copies of the Software, and to
9  * permit persons to whom the Software is furnished to do so, subject to
10  * the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be
13  * included in all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
16  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
17  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
18  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
19  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
20  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
21  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "inner.h"
26 
27 /*
28  * In this file, we handle big integers with a custom format, i.e.
29  * without the usual one-word header. Value is split into 31-bit words,
30  * each stored in a 32-bit slot (top bit is zero) in little-endian
31  * order. The length (in words) is provided explicitly. In some cases,
32  * the value can be negative (using two's complement representation). In
33  * some cases, the top word is allowed to have a 32th bit.
34  */
35 
36 /*
37  * Negate big integer conditionally. The value consists of 'len' words,
38  * with 31 bits in each word (the top bit of each word should be 0,
39  * except possibly for the last word). If 'ctl' is 1, the negation is
40  * computed; otherwise, if 'ctl' is 0, then the value is unchanged.
41  */
42 static void
43 cond_negate(uint32_t *a, size_t len, uint32_t ctl)
44 {
45 	size_t k;
46 	uint32_t cc, xm;
47 
48 	cc = ctl;
49 	xm = -ctl >> 1;
50 	for (k = 0; k < len; k ++) {
51 		uint32_t aw;
52 
53 		aw = a[k];
54 		aw = (aw ^ xm) + cc;
55 		a[k] = aw & 0x7FFFFFFF;
56 		cc = aw >> 31;
57 	}
58 }
59 
60 /*
61  * Finish modular reduction. Rules on input parameters:
62  *
63  *   if neg = 1, then -m <= a < 0
64  *   if neg = 0, then 0 <= a < 2*m
65  *
66  * If neg = 0, then the top word of a[] may use 32 bits.
67  *
68  * Also, modulus m must be odd.
69  */
70 static void
71 finish_mod(uint32_t *a, size_t len, const uint32_t *m, uint32_t neg)
72 {
73 	size_t k;
74 	uint32_t cc, xm, ym;
75 
76 	/*
77 	 * First pass: compare a (assumed nonnegative) with m.
78 	 * Note that if the final word uses the top extra bit, then
79 	 * subtracting m must yield a value less than 2^31, since we
80 	 * assumed that a < 2*m.
81 	 */
82 	cc = 0;
83 	for (k = 0; k < len; k ++) {
84 		uint32_t aw, mw;
85 
86 		aw = a[k];
87 		mw = m[k];
88 		cc = (aw - mw - cc) >> 31;
89 	}
90 
91 	/*
92 	 * At this point:
93 	 *   if neg = 1, then we must add m (regardless of cc)
94 	 *   if neg = 0 and cc = 0, then we must subtract m
95 	 *   if neg = 0 and cc = 1, then we must do nothing
96 	 */
97 	xm = -neg >> 1;
98 	ym = -(neg | (1 - cc));
99 	cc = neg;
100 	for (k = 0; k < len; k ++) {
101 		uint32_t aw, mw;
102 
103 		aw = a[k];
104 		mw = (m[k] ^ xm) & ym;
105 		aw = aw - mw - cc;
106 		a[k] = aw & 0x7FFFFFFF;
107 		cc = aw >> 31;
108 	}
109 }
110 
111 /*
112  * Compute:
113  *   a <- (a*pa+b*pb)/(2^31)
114  *   b <- (a*qa+b*qb)/(2^31)
115  * The division is assumed to be exact (i.e. the low word is dropped).
116  * If the final a is negative, then it is negated. Similarly for b.
117  * Returned value is the combination of two bits:
118  *   bit 0: 1 if a had to be negated, 0 otherwise
119  *   bit 1: 1 if b had to be negated, 0 otherwise
120  *
121  * Factors pa, pb, qa and qb must be at most 2^31 in absolute value.
122  * Source integers a and b must be nonnegative; top word is not allowed
123  * to contain an extra 32th bit.
124  */
125 static uint32_t
126 co_reduce(uint32_t *a, uint32_t *b, size_t len,
127 	int64_t pa, int64_t pb, int64_t qa, int64_t qb)
128 {
129 	size_t k;
130 	int64_t cca, ccb;
131 	uint32_t nega, negb;
132 
133 	cca = 0;
134 	ccb = 0;
135 	for (k = 0; k < len; k ++) {
136 		uint32_t wa, wb;
137 		uint64_t za, zb;
138 		uint64_t tta, ttb;
139 
140 		/*
141 		 * Since:
142 		 *   |pa| <= 2^31
143 		 *   |pb| <= 2^31
144 		 *   0 <= wa <= 2^31 - 1
145 		 *   0 <= wb <= 2^31 - 1
146 		 *   |cca| <= 2^32 - 1
147 		 * Then:
148 		 *   |za| <= (2^31-1)*(2^32) + (2^32-1) = 2^63 - 1
149 		 *
150 		 * Thus, the new value of cca is such that |cca| <= 2^32 - 1.
151 		 * The same applies to ccb.
152 		 */
153 		wa = a[k];
154 		wb = b[k];
155 		za = wa * (uint64_t)pa + wb * (uint64_t)pb + (uint64_t)cca;
156 		zb = wa * (uint64_t)qa + wb * (uint64_t)qb + (uint64_t)ccb;
157 		if (k > 0) {
158 			a[k - 1] = za & 0x7FFFFFFF;
159 			b[k - 1] = zb & 0x7FFFFFFF;
160 		}
161 
162 		/*
163 		 * For the new values of cca and ccb, we need a signed
164 		 * right-shift; since, in C, right-shifting a signed
165 		 * negative value is implementation-defined, we use a
166 		 * custom portable sign extension expression.
167 		 */
168 #define M   ((uint64_t)1 << 32)
169 		tta = za >> 31;
170 		ttb = zb >> 31;
171 		tta = (tta ^ M) - M;
172 		ttb = (ttb ^ M) - M;
173 		cca = *(int64_t *)&tta;
174 		ccb = *(int64_t *)&ttb;
175 #undef M
176 	}
177 	a[len - 1] = (uint32_t)cca;
178 	b[len - 1] = (uint32_t)ccb;
179 
180 	nega = (uint32_t)((uint64_t)cca >> 63);
181 	negb = (uint32_t)((uint64_t)ccb >> 63);
182 	cond_negate(a, len, nega);
183 	cond_negate(b, len, negb);
184 	return nega | (negb << 1);
185 }
186 
187 /*
188  * Compute:
189  *   a <- (a*pa+b*pb)/(2^31) mod m
190  *   b <- (a*qa+b*qb)/(2^31) mod m
191  *
192  * m0i is equal to -1/m[0] mod 2^31.
193  *
194  * Factors pa, pb, qa and qb must be at most 2^31 in absolute value.
195  * Source integers a and b must be nonnegative; top word is not allowed
196  * to contain an extra 32th bit.
197  */
198 static void
199 co_reduce_mod(uint32_t *a, uint32_t *b, size_t len,
200 	int64_t pa, int64_t pb, int64_t qa, int64_t qb,
201 	const uint32_t *m, uint32_t m0i)
202 {
203 	size_t k;
204 	int64_t cca, ccb;
205 	uint32_t fa, fb;
206 
207 	cca = 0;
208 	ccb = 0;
209 	fa = ((a[0] * (uint32_t)pa + b[0] * (uint32_t)pb) * m0i) & 0x7FFFFFFF;
210 	fb = ((a[0] * (uint32_t)qa + b[0] * (uint32_t)qb) * m0i) & 0x7FFFFFFF;
211 	for (k = 0; k < len; k ++) {
212 		uint32_t wa, wb;
213 		uint64_t za, zb;
214 		uint64_t tta, ttb;
215 
216 		/*
217 		 * In this loop, carries 'cca' and 'ccb' always fit on
218 		 * 33 bits (in absolute value).
219 		 */
220 		wa = a[k];
221 		wb = b[k];
222 		za = wa * (uint64_t)pa + wb * (uint64_t)pb
223 			+ m[k] * (uint64_t)fa + (uint64_t)cca;
224 		zb = wa * (uint64_t)qa + wb * (uint64_t)qb
225 			+ m[k] * (uint64_t)fb + (uint64_t)ccb;
226 		if (k > 0) {
227 			a[k - 1] = (uint32_t)za & 0x7FFFFFFF;
228 			b[k - 1] = (uint32_t)zb & 0x7FFFFFFF;
229 		}
230 
231 #define M   ((uint64_t)1 << 32)
232 		tta = za >> 31;
233 		ttb = zb >> 31;
234 		tta = (tta ^ M) - M;
235 		ttb = (ttb ^ M) - M;
236 		cca = *(int64_t *)&tta;
237 		ccb = *(int64_t *)&ttb;
238 #undef M
239 	}
240 	a[len - 1] = (uint32_t)cca;
241 	b[len - 1] = (uint32_t)ccb;
242 
243 	/*
244 	 * At this point:
245 	 *   -m <= a < 2*m
246 	 *   -m <= b < 2*m
247 	 * (this is a case of Montgomery reduction)
248 	 * The top word of 'a' and 'b' may have a 32-th bit set.
249 	 * We may have to add or subtract the modulus.
250 	 */
251 	finish_mod(a, len, m, (uint32_t)((uint64_t)cca >> 63));
252 	finish_mod(b, len, m, (uint32_t)((uint64_t)ccb >> 63));
253 }
254 
255 /* see inner.h */
256 uint32_t
257 br_i31_moddiv(uint32_t *x, const uint32_t *y, const uint32_t *m, uint32_t m0i,
258 	uint32_t *t)
259 {
260 	/*
261 	 * Algorithm is an extended binary GCD. We maintain four values
262 	 * a, b, u and v, with the following invariants:
263 	 *
264 	 *   a * x = y * u mod m
265 	 *   b * x = y * v mod m
266 	 *
267 	 * Starting values are:
268 	 *
269 	 *   a = y
270 	 *   b = m
271 	 *   u = x
272 	 *   v = 0
273 	 *
274 	 * The formal definition of the algorithm is a sequence of steps:
275 	 *
276 	 *   - If a is even, then a <- a/2 and u <- u/2 mod m.
277 	 *   - Otherwise, if b is even, then b <- b/2 and v <- v/2 mod m.
278 	 *   - Otherwise, if a > b, then a <- (a-b)/2 and u <- (u-v)/2 mod m.
279 	 *   - Otherwise, b <- (b-a)/2 and v <- (v-u)/2 mod m.
280 	 *
281 	 * Algorithm stops when a = b. At that point, they both are equal
282 	 * to GCD(y,m); the modular division succeeds if that value is 1.
283 	 * The result of the modular division is then u (or v: both are
284 	 * equal at that point).
285 	 *
286 	 * Each step makes either a or b shrink by at least one bit; hence,
287 	 * if m has bit length k bits, then 2k-2 steps are sufficient.
288 	 *
289 	 *
290 	 * Though complexity is quadratic in the size of m, the bit-by-bit
291 	 * processing is not very efficient. We can speed up processing by
292 	 * remarking that the decisions are taken based only on observation
293 	 * of the top and low bits of a and b.
294 	 *
295 	 * In the loop below, at each iteration, we use the two top words
296 	 * of a and b, and the low words of a and b, to compute reduction
297 	 * parameters pa, pb, qa and qb such that the new values for a
298 	 * and b are:
299 	 *
300 	 *   a' = (a*pa + b*pb) / (2^31)
301 	 *   b' = (a*qa + b*qb) / (2^31)
302 	 *
303 	 * the division being exact.
304 	 *
305 	 * Since the choices are based on the top words, they may be slightly
306 	 * off, requiring an optional correction: if a' < 0, then we replace
307 	 * pa with -pa, and pb with -pb. The total length of a and b is
308 	 * thus reduced by at least 30 bits at each iteration.
309 	 *
310 	 * The stopping conditions are still the same, though: when a
311 	 * and b become equal, they must be both odd (since m is odd,
312 	 * the GCD cannot be even), therefore the next operation is a
313 	 * subtraction, and one of the values becomes 0. At that point,
314 	 * nothing else happens, i.e. one value is stuck at 0, and the
315 	 * other one is the GCD.
316 	 */
317 	size_t len, k;
318 	uint32_t *a, *b, *u, *v;
319 	uint32_t num, r;
320 
321 	len = (m[0] + 31) >> 5;
322 	a = t;
323 	b = a + len;
324 	u = x + 1;
325 	v = b + len;
326 	memcpy(a, y + 1, len * sizeof *y);
327 	memcpy(b, m + 1, len * sizeof *m);
328 	memset(v, 0, len * sizeof *v);
329 
330 	/*
331 	 * Loop below ensures that a and b are reduced by some bits each,
332 	 * for a total of at least 30 bits.
333 	 */
334 	for (num = ((m[0] - (m[0] >> 5)) << 1) + 30; num >= 30; num -= 30) {
335 		size_t j;
336 		uint32_t c0, c1;
337 		uint32_t a0, a1, b0, b1;
338 		uint64_t a_hi, b_hi;
339 		uint32_t a_lo, b_lo;
340 		int64_t pa, pb, qa, qb;
341 		int i;
342 
343 		/*
344 		 * Extract top words of a and b. If j is the highest
345 		 * index >= 1 such that a[j] != 0 or b[j] != 0, then we want
346 		 * (a[j] << 31) + a[j - 1], and (b[j] << 31) + b[j - 1].
347 		 * If a and b are down to one word each, then we use a[0]
348 		 * and b[0].
349 		 */
350 		c0 = (uint32_t)-1;
351 		c1 = (uint32_t)-1;
352 		a0 = 0;
353 		a1 = 0;
354 		b0 = 0;
355 		b1 = 0;
356 		j = len;
357 		while (j -- > 0) {
358 			uint32_t aw, bw;
359 
360 			aw = a[j];
361 			bw = b[j];
362 			a0 ^= (a0 ^ aw) & c0;
363 			a1 ^= (a1 ^ aw) & c1;
364 			b0 ^= (b0 ^ bw) & c0;
365 			b1 ^= (b1 ^ bw) & c1;
366 			c1 = c0;
367 			c0 &= (((aw | bw) + 0x7FFFFFFF) >> 31) - (uint32_t)1;
368 		}
369 
370 		/*
371 		 * If c1 = 0, then we grabbed two words for a and b.
372 		 * If c1 != 0 but c0 = 0, then we grabbed one word. It
373 		 * is not possible that c1 != 0 and c0 != 0, because that
374 		 * would mean that both integers are zero.
375 		 */
376 		a1 |= a0 & c1;
377 		a0 &= ~c1;
378 		b1 |= b0 & c1;
379 		b0 &= ~c1;
380 		a_hi = ((uint64_t)a0 << 31) + a1;
381 		b_hi = ((uint64_t)b0 << 31) + b1;
382 		a_lo = a[0];
383 		b_lo = b[0];
384 
385 		/*
386 		 * Compute reduction factors:
387 		 *
388 		 *   a' = a*pa + b*pb
389 		 *   b' = a*qa + b*qb
390 		 *
391 		 * such that a' and b' are both multiple of 2^31, but are
392 		 * only marginally larger than a and b.
393 		 */
394 		pa = 1;
395 		pb = 0;
396 		qa = 0;
397 		qb = 1;
398 		for (i = 0; i < 31; i ++) {
399 			/*
400 			 * At each iteration:
401 			 *
402 			 *   a <- (a-b)/2 if: a is odd, b is odd, a_hi > b_hi
403 			 *   b <- (b-a)/2 if: a is odd, b is odd, a_hi <= b_hi
404 			 *   a <- a/2 if: a is even
405 			 *   b <- b/2 if: a is odd, b is even
406 			 *
407 			 * We multiply a_lo and b_lo by 2 at each
408 			 * iteration, thus a division by 2 really is a
409 			 * non-multiplication by 2.
410 			 */
411 			uint32_t r, oa, ob, cAB, cBA, cA;
412 			uint64_t rz;
413 
414 			/*
415 			 * r = GT(a_hi, b_hi)
416 			 * But the GT() function works on uint32_t operands,
417 			 * so we inline a 64-bit version here.
418 			 */
419 			rz = b_hi - a_hi;
420 			r = (uint32_t)((rz ^ ((a_hi ^ b_hi)
421 				& (a_hi ^ rz))) >> 63);
422 
423 			/*
424 			 * cAB = 1 if b must be subtracted from a
425 			 * cBA = 1 if a must be subtracted from b
426 			 * cA = 1 if a is divided by 2, 0 otherwise
427 			 *
428 			 * Rules:
429 			 *
430 			 *   cAB and cBA cannot be both 1.
431 			 *   if a is not divided by 2, b is.
432 			 */
433 			oa = (a_lo >> i) & 1;
434 			ob = (b_lo >> i) & 1;
435 			cAB = oa & ob & r;
436 			cBA = oa & ob & NOT(r);
437 			cA = cAB | NOT(oa);
438 
439 			/*
440 			 * Conditional subtractions.
441 			 */
442 			a_lo -= b_lo & -cAB;
443 			a_hi -= b_hi & -(uint64_t)cAB;
444 			pa -= qa & -(int64_t)cAB;
445 			pb -= qb & -(int64_t)cAB;
446 			b_lo -= a_lo & -cBA;
447 			b_hi -= a_hi & -(uint64_t)cBA;
448 			qa -= pa & -(int64_t)cBA;
449 			qb -= pb & -(int64_t)cBA;
450 
451 			/*
452 			 * Shifting.
453 			 */
454 			a_lo += a_lo & (cA - 1);
455 			pa += pa & ((int64_t)cA - 1);
456 			pb += pb & ((int64_t)cA - 1);
457 			a_hi ^= (a_hi ^ (a_hi >> 1)) & -(uint64_t)cA;
458 			b_lo += b_lo & -cA;
459 			qa += qa & -(int64_t)cA;
460 			qb += qb & -(int64_t)cA;
461 			b_hi ^= (b_hi ^ (b_hi >> 1)) & ((uint64_t)cA - 1);
462 		}
463 
464 		/*
465 		 * Replace a and b with new values a' and b'.
466 		 */
467 		r = co_reduce(a, b, len, pa, pb, qa, qb);
468 		pa -= pa * ((r & 1) << 1);
469 		pb -= pb * ((r & 1) << 1);
470 		qa -= qa * (r & 2);
471 		qb -= qb * (r & 2);
472 		co_reduce_mod(u, v, len, pa, pb, qa, qb, m + 1, m0i);
473 	}
474 
475 	/*
476 	 * Now one of the arrays should be 0, and the other contains
477 	 * the GCD. If a is 0, then u is 0 as well, and v contains
478 	 * the division result.
479 	 * Result is correct if and only if GCD is 1.
480 	 */
481 	r = (a[0] | b[0]) ^ 1;
482 	u[0] |= v[0];
483 	for (k = 1; k < len; k ++) {
484 		r |= a[k] | b[k];
485 		u[k] |= v[k];
486 	}
487 	return EQ0(r);
488 }
489