xref: /freebsd/contrib/bearssl/src/int/i62_modpow2.c (revision f126890a)
1 /*
2  * Copyright (c) 2017 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 #if BR_INT128 || BR_UMUL128
28 
29 #if BR_INT128
30 
31 /*
32  * Compute x*y+v1+v2. Operands are 64-bit, and result is 128-bit, with
33  * high word in "hi" and low word in "lo".
34  */
35 #define FMA1(hi, lo, x, y, v1, v2)   do { \
36 		unsigned __int128 fmaz; \
37 		fmaz = (unsigned __int128)(x) * (unsigned __int128)(y) \
38 			+ (unsigned __int128)(v1) + (unsigned __int128)(v2); \
39 		(hi) = (uint64_t)(fmaz >> 64); \
40 		(lo) = (uint64_t)fmaz; \
41 	} while (0)
42 
43 /*
44  * Compute x1*y1+x2*y2+v1+v2. Operands are 64-bit, and result is 128-bit,
45  * with high word in "hi" and low word in "lo".
46  *
47  * Callers should ensure that the two inner products, and the v1 and v2
48  * operands, are multiple of 4 (this is not used by this specific definition
49  * but may help other implementations).
50  */
51 #define FMA2(hi, lo, x1, y1, x2, y2, v1, v2)   do { \
52 		unsigned __int128 fmaz; \
53 		fmaz = (unsigned __int128)(x1) * (unsigned __int128)(y1) \
54 			+ (unsigned __int128)(x2) * (unsigned __int128)(y2) \
55 			+ (unsigned __int128)(v1) + (unsigned __int128)(v2); \
56 		(hi) = (uint64_t)(fmaz >> 64); \
57 		(lo) = (uint64_t)fmaz; \
58 	} while (0)
59 
60 #elif BR_UMUL128
61 
62 #include <intrin.h>
63 
64 #define FMA1(hi, lo, x, y, v1, v2)   do { \
65 		uint64_t fmahi, fmalo; \
66 		unsigned char fmacc; \
67 		fmalo = _umul128((x), (y), &fmahi); \
68 		fmacc = _addcarry_u64(0, fmalo, (v1), &fmalo); \
69 		_addcarry_u64(fmacc, fmahi, 0, &fmahi); \
70 		fmacc = _addcarry_u64(0, fmalo, (v2), &(lo)); \
71 		_addcarry_u64(fmacc, fmahi, 0, &(hi)); \
72 	} while (0)
73 
74 /*
75  * Normally we should use _addcarry_u64() for FMA2 too, but it makes
76  * Visual Studio crash. Instead we use this version, which leverages
77  * the fact that the vx operands, and the products, are multiple of 4.
78  * This is unfortunately slower.
79  */
80 #define FMA2(hi, lo, x1, y1, x2, y2, v1, v2)   do { \
81 		uint64_t fma1hi, fma1lo; \
82 		uint64_t fma2hi, fma2lo; \
83 		uint64_t fmatt; \
84 		fma1lo = _umul128((x1), (y1), &fma1hi); \
85 		fma2lo = _umul128((x2), (y2), &fma2hi); \
86 		fmatt = (fma1lo >> 2) + (fma2lo >> 2) \
87 			+ ((v1) >> 2) + ((v2) >> 2); \
88 		(lo) = fmatt << 2; \
89 		(hi) = fma1hi + fma2hi + (fmatt >> 62); \
90 	} while (0)
91 
92 /*
93  * The FMA2 macro definition we would prefer to use, but it triggers
94  * an internal compiler error in Visual Studio 2015.
95  *
96 #define FMA2(hi, lo, x1, y1, x2, y2, v1, v2)   do { \
97 		uint64_t fma1hi, fma1lo; \
98 		uint64_t fma2hi, fma2lo; \
99 		unsigned char fmacc; \
100 		fma1lo = _umul128((x1), (y1), &fma1hi); \
101 		fma2lo = _umul128((x2), (y2), &fma2hi); \
102 		fmacc = _addcarry_u64(0, fma1lo, (v1), &fma1lo); \
103 		_addcarry_u64(fmacc, fma1hi, 0, &fma1hi); \
104 		fmacc = _addcarry_u64(0, fma2lo, (v2), &fma2lo); \
105 		_addcarry_u64(fmacc, fma2hi, 0, &fma2hi); \
106 		fmacc = _addcarry_u64(0, fma1lo, fma2lo, &(lo)); \
107 		_addcarry_u64(fmacc, fma1hi, fma2hi, &(hi)); \
108 	} while (0)
109  */
110 
111 #endif
112 
113 #define MASK62           ((uint64_t)0x3FFFFFFFFFFFFFFF)
114 #define MUL62_lo(x, y)   (((uint64_t)(x) * (uint64_t)(y)) & MASK62)
115 
116 /*
117  * Subtract b from a, and return the final carry. If 'ctl32' is 0, then
118  * a[] is kept unmodified, but the final carry is still computed and
119  * returned.
120  */
121 static uint32_t
122 i62_sub(uint64_t *a, const uint64_t *b, size_t num, uint32_t ctl32)
123 {
124 	uint64_t cc, mask;
125 	size_t u;
126 
127 	cc = 0;
128 	ctl32 = -ctl32;
129 	mask = (uint64_t)ctl32 | ((uint64_t)ctl32 << 32);
130 	for (u = 0; u < num; u ++) {
131 		uint64_t aw, bw, dw;
132 
133 		aw = a[u];
134 		bw = b[u];
135 		dw = aw - bw - cc;
136 		cc = dw >> 63;
137 		dw &= MASK62;
138 		a[u] = aw ^ (mask & (dw ^ aw));
139 	}
140 	return (uint32_t)cc;
141 }
142 
143 /*
144  * Montgomery multiplication, over arrays of 62-bit values. The
145  * destination array (d) must be distinct from the other operands
146  * (x, y and m). All arrays are in little-endian format (least
147  * significant word comes first) over 'num' words.
148  */
149 static void
150 montymul(uint64_t *d, const uint64_t *x, const uint64_t *y,
151 	const uint64_t *m, size_t num, uint64_t m0i)
152 {
153 	uint64_t dh;
154 	size_t u, num4;
155 
156 	num4 = 1 + ((num - 1) & ~(size_t)3);
157 	memset(d, 0, num * sizeof *d);
158 	dh = 0;
159 	for (u = 0; u < num; u ++) {
160 		size_t v;
161 		uint64_t f, xu;
162 		uint64_t r, zh;
163 		uint64_t hi, lo;
164 
165 		xu = x[u] << 2;
166 		f = MUL62_lo(d[0] + MUL62_lo(x[u], y[0]), m0i) << 2;
167 
168 		FMA2(hi, lo, xu, y[0], f, m[0], d[0] << 2, 0);
169 		r = hi;
170 
171 		for (v = 1; v < num4; v += 4) {
172 			FMA2(hi, lo, xu, y[v + 0],
173 				f, m[v + 0], d[v + 0] << 2, r << 2);
174 			r = hi + (r >> 62);
175 			d[v - 1] = lo >> 2;
176 			FMA2(hi, lo, xu, y[v + 1],
177 				f, m[v + 1], d[v + 1] << 2, r << 2);
178 			r = hi + (r >> 62);
179 			d[v + 0] = lo >> 2;
180 			FMA2(hi, lo, xu, y[v + 2],
181 				f, m[v + 2], d[v + 2] << 2, r << 2);
182 			r = hi + (r >> 62);
183 			d[v + 1] = lo >> 2;
184 			FMA2(hi, lo, xu, y[v + 3],
185 				f, m[v + 3], d[v + 3] << 2, r << 2);
186 			r = hi + (r >> 62);
187 			d[v + 2] = lo >> 2;
188 		}
189 		for (; v < num; v ++) {
190 			FMA2(hi, lo, xu, y[v], f, m[v], d[v] << 2, r << 2);
191 			r = hi + (r >> 62);
192 			d[v - 1] = lo >> 2;
193 		}
194 
195 		zh = dh + r;
196 		d[num - 1] = zh & MASK62;
197 		dh = zh >> 62;
198 	}
199 	i62_sub(d, m, num, (uint32_t)dh | NOT(i62_sub(d, m, num, 0)));
200 }
201 
202 /*
203  * Conversion back from Montgomery representation.
204  */
205 static void
206 frommonty(uint64_t *x, const uint64_t *m, size_t num, uint64_t m0i)
207 {
208 	size_t u, v;
209 
210 	for (u = 0; u < num; u ++) {
211 		uint64_t f, cc;
212 
213 		f = MUL62_lo(x[0], m0i) << 2;
214 		cc = 0;
215 		for (v = 0; v < num; v ++) {
216 			uint64_t hi, lo;
217 
218 			FMA1(hi, lo, f, m[v], x[v] << 2, cc);
219 			cc = hi << 2;
220 			if (v != 0) {
221 				x[v - 1] = lo >> 2;
222 			}
223 		}
224 		x[num - 1] = cc >> 2;
225 	}
226 	i62_sub(x, m, num, NOT(i62_sub(x, m, num, 0)));
227 }
228 
229 /* see inner.h */
230 uint32_t
231 br_i62_modpow_opt(uint32_t *x31, const unsigned char *e, size_t elen,
232 	const uint32_t *m31, uint32_t m0i31, uint64_t *tmp, size_t twlen)
233 {
234 	size_t u, mw31num, mw62num;
235 	uint64_t *x, *m, *t1, *t2;
236 	uint64_t m0i;
237 	uint32_t acc;
238 	int win_len, acc_len;
239 
240 	/*
241 	 * Get modulus size, in words.
242 	 */
243 	mw31num = (m31[0] + 31) >> 5;
244 	mw62num = (mw31num + 1) >> 1;
245 
246 	/*
247 	 * In order to apply this function, we must have enough room to
248 	 * copy the operand and modulus into the temporary array, along
249 	 * with at least two temporaries. If there is not enough room,
250 	 * switch to br_i31_modpow(). We also use br_i31_modpow() if the
251 	 * modulus length is not at least four words (94 bits or more).
252 	 */
253 	if (mw31num < 4 || (mw62num << 2) > twlen) {
254 		/*
255 		 * We assume here that we can split an aligned uint64_t
256 		 * into two properly aligned uint32_t. Since both types
257 		 * are supposed to have an exact width with no padding,
258 		 * then this property must hold.
259 		 */
260 		size_t txlen;
261 
262 		txlen = mw31num + 1;
263 		if (twlen < txlen) {
264 			return 0;
265 		}
266 		br_i31_modpow(x31, e, elen, m31, m0i31,
267 			(uint32_t *)tmp, (uint32_t *)tmp + txlen);
268 		return 1;
269 	}
270 
271 	/*
272 	 * Convert x to Montgomery representation: this means that
273 	 * we replace x with x*2^z mod m, where z is the smallest multiple
274 	 * of the word size such that 2^z >= m. We want to reuse the 31-bit
275 	 * functions here (for constant-time operation), but we need z
276 	 * for a 62-bit word size.
277 	 */
278 	for (u = 0; u < mw62num; u ++) {
279 		br_i31_muladd_small(x31, 0, m31);
280 		br_i31_muladd_small(x31, 0, m31);
281 	}
282 
283 	/*
284 	 * Assemble operands into arrays of 62-bit words. Note that
285 	 * all the arrays of 62-bit words that we will handle here
286 	 * are without any leading size word.
287 	 *
288 	 * We also adjust tmp and twlen to account for the words used
289 	 * for these extra arrays.
290 	 */
291 	m = tmp;
292 	x = tmp + mw62num;
293 	tmp += (mw62num << 1);
294 	twlen -= (mw62num << 1);
295 	for (u = 0; u < mw31num; u += 2) {
296 		size_t v;
297 
298 		v = u >> 1;
299 		if ((u + 1) == mw31num) {
300 			m[v] = (uint64_t)m31[u + 1];
301 			x[v] = (uint64_t)x31[u + 1];
302 		} else {
303 			m[v] = (uint64_t)m31[u + 1]
304 				+ ((uint64_t)m31[u + 2] << 31);
305 			x[v] = (uint64_t)x31[u + 1]
306 				+ ((uint64_t)x31[u + 2] << 31);
307 		}
308 	}
309 
310 	/*
311 	 * Compute window size. We support windows up to 5 bits; for a
312 	 * window of size k bits, we need 2^k+1 temporaries (for k = 1,
313 	 * we use special code that uses only 2 temporaries).
314 	 */
315 	for (win_len = 5; win_len > 1; win_len --) {
316 		if ((((uint32_t)1 << win_len) + 1) * mw62num <= twlen) {
317 			break;
318 		}
319 	}
320 
321 	t1 = tmp;
322 	t2 = tmp + mw62num;
323 
324 	/*
325 	 * Compute m0i, which is equal to -(1/m0) mod 2^62. We were
326 	 * provided with m0i31, which already fulfills this property
327 	 * modulo 2^31; the single expression below is then sufficient.
328 	 */
329 	m0i = (uint64_t)m0i31;
330 	m0i = MUL62_lo(m0i, (uint64_t)2 + MUL62_lo(m0i, m[0]));
331 
332 	/*
333 	 * Compute window contents. If the window has size one bit only,
334 	 * then t2 is set to x; otherwise, t2[0] is left untouched, and
335 	 * t2[k] is set to x^k (for k >= 1).
336 	 */
337 	if (win_len == 1) {
338 		memcpy(t2, x, mw62num * sizeof *x);
339 	} else {
340 		uint64_t *base;
341 
342 		memcpy(t2 + mw62num, x, mw62num * sizeof *x);
343 		base = t2 + mw62num;
344 		for (u = 2; u < ((unsigned)1 << win_len); u ++) {
345 			montymul(base + mw62num, base, x, m, mw62num, m0i);
346 			base += mw62num;
347 		}
348 	}
349 
350 	/*
351 	 * Set x to 1, in Montgomery representation. We again use the
352 	 * 31-bit code.
353 	 */
354 	br_i31_zero(x31, m31[0]);
355 	x31[(m31[0] + 31) >> 5] = 1;
356 	br_i31_muladd_small(x31, 0, m31);
357 	if (mw31num & 1) {
358 		br_i31_muladd_small(x31, 0, m31);
359 	}
360 	for (u = 0; u < mw31num; u += 2) {
361 		size_t v;
362 
363 		v = u >> 1;
364 		if ((u + 1) == mw31num) {
365 			x[v] = (uint64_t)x31[u + 1];
366 		} else {
367 			x[v] = (uint64_t)x31[u + 1]
368 				+ ((uint64_t)x31[u + 2] << 31);
369 		}
370 	}
371 
372 	/*
373 	 * We process bits from most to least significant. At each
374 	 * loop iteration, we have acc_len bits in acc.
375 	 */
376 	acc = 0;
377 	acc_len = 0;
378 	while (acc_len > 0 || elen > 0) {
379 		int i, k;
380 		uint32_t bits;
381 		uint64_t mask1, mask2;
382 
383 		/*
384 		 * Get the next bits.
385 		 */
386 		k = win_len;
387 		if (acc_len < win_len) {
388 			if (elen > 0) {
389 				acc = (acc << 8) | *e ++;
390 				elen --;
391 				acc_len += 8;
392 			} else {
393 				k = acc_len;
394 			}
395 		}
396 		bits = (acc >> (acc_len - k)) & (((uint32_t)1 << k) - 1);
397 		acc_len -= k;
398 
399 		/*
400 		 * We could get exactly k bits. Compute k squarings.
401 		 */
402 		for (i = 0; i < k; i ++) {
403 			montymul(t1, x, x, m, mw62num, m0i);
404 			memcpy(x, t1, mw62num * sizeof *x);
405 		}
406 
407 		/*
408 		 * Window lookup: we want to set t2 to the window
409 		 * lookup value, assuming the bits are non-zero. If
410 		 * the window length is 1 bit only, then t2 is
411 		 * already set; otherwise, we do a constant-time lookup.
412 		 */
413 		if (win_len > 1) {
414 			uint64_t *base;
415 
416 			memset(t2, 0, mw62num * sizeof *t2);
417 			base = t2 + mw62num;
418 			for (u = 1; u < ((uint32_t)1 << k); u ++) {
419 				uint64_t mask;
420 				size_t v;
421 
422 				mask = -(uint64_t)EQ(u, bits);
423 				for (v = 0; v < mw62num; v ++) {
424 					t2[v] |= mask & base[v];
425 				}
426 				base += mw62num;
427 			}
428 		}
429 
430 		/*
431 		 * Multiply with the looked-up value. We keep the product
432 		 * only if the exponent bits are not all-zero.
433 		 */
434 		montymul(t1, x, t2, m, mw62num, m0i);
435 		mask1 = -(uint64_t)EQ(bits, 0);
436 		mask2 = ~mask1;
437 		for (u = 0; u < mw62num; u ++) {
438 			x[u] = (mask1 & x[u]) | (mask2 & t1[u]);
439 		}
440 	}
441 
442 	/*
443 	 * Convert back from Montgomery representation.
444 	 */
445 	frommonty(x, m, mw62num, m0i);
446 
447 	/*
448 	 * Convert result into 31-bit words.
449 	 */
450 	for (u = 0; u < mw31num; u += 2) {
451 		uint64_t zw;
452 
453 		zw = x[u >> 1];
454 		x31[u + 1] = (uint32_t)zw & 0x7FFFFFFF;
455 		if ((u + 1) < mw31num) {
456 			x31[u + 2] = (uint32_t)(zw >> 31);
457 		}
458 	}
459 	return 1;
460 }
461 
462 #else
463 
464 /* see inner.h */
465 uint32_t
466 br_i62_modpow_opt(uint32_t *x31, const unsigned char *e, size_t elen,
467 	const uint32_t *m31, uint32_t m0i31, uint64_t *tmp, size_t twlen)
468 {
469 	size_t mwlen;
470 
471 	mwlen = (m31[0] + 63) >> 5;
472 	if (twlen < mwlen) {
473 		return 0;
474 	}
475 	return br_i31_modpow_opt(x31, e, elen, m31, m0i31,
476 		(uint32_t *)tmp, twlen << 1);
477 }
478 
479 #endif
480 
481 /* see inner.h */
482 uint32_t
483 br_i62_modpow_opt_as_i31(uint32_t *x31, const unsigned char *e, size_t elen,
484 	const uint32_t *m31, uint32_t m0i31, uint32_t *tmp, size_t twlen)
485 {
486 	/*
487 	 * As documented, this function expects the 'tmp' argument to be
488 	 * 64-bit aligned. This is OK since this function is internal (it
489 	 * is not part of BearSSL's public API).
490 	 */
491 	return br_i62_modpow_opt(x31, e, elen, m31, m0i31,
492 		(uint64_t *)tmp, twlen >> 1);
493 }
494