xref: /openbmc/linux/lib/crypto/mpi/mpih-div.c (revision 86e281fc)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* mpihelp-div.c  -  MPI helper functions
3  *	Copyright (C) 1994, 1996 Free Software Foundation, Inc.
4  *	Copyright (C) 1998, 1999 Free Software Foundation, Inc.
5  *
6  * This file is part of GnuPG.
7  *
8  * Note: This code is heavily based on the GNU MP Library.
9  *	 Actually it's the same code with only minor changes in the
10  *	 way the data is stored; this is to support the abstraction
11  *	 of an optional secure memory allocation which may be used
12  *	 to avoid revealing of sensitive data due to paging etc.
13  *	 The GNU MP Library itself is published under the LGPL;
14  *	 however I decided to publish this code under the plain GPL.
15  */
16 
17 #include "mpi-internal.h"
18 #include "longlong.h"
19 
20 #ifndef UMUL_TIME
21 #define UMUL_TIME 1
22 #endif
23 #ifndef UDIV_TIME
24 #define UDIV_TIME UMUL_TIME
25 #endif
26 
27 
28 mpi_limb_t
29 mpihelp_mod_1(mpi_ptr_t dividend_ptr, mpi_size_t dividend_size,
30 			mpi_limb_t divisor_limb)
31 {
32 	mpi_size_t i;
33 	mpi_limb_t n1, n0, r;
34 	mpi_limb_t dummy __maybe_unused;
35 
36 	/* Botch: Should this be handled at all?  Rely on callers?	*/
37 	if (!dividend_size)
38 		return 0;
39 
40 	/* If multiplication is much faster than division, and the
41 	 * dividend is large, pre-invert the divisor, and use
42 	 * only multiplications in the inner loop.
43 	 *
44 	 * This test should be read:
45 	 *	 Does it ever help to use udiv_qrnnd_preinv?
46 	 *	   && Does what we save compensate for the inversion overhead?
47 	 */
48 	if (UDIV_TIME > (2 * UMUL_TIME + 6)
49 			&& (UDIV_TIME - (2 * UMUL_TIME + 6)) * dividend_size > UDIV_TIME) {
50 		int normalization_steps;
51 
52 		normalization_steps = count_leading_zeros(divisor_limb);
53 		if (normalization_steps) {
54 			mpi_limb_t divisor_limb_inverted;
55 
56 			divisor_limb <<= normalization_steps;
57 
58 			/* Compute (2**2N - 2**N * DIVISOR_LIMB) / DIVISOR_LIMB.  The
59 			 * result is a (N+1)-bit approximation to 1/DIVISOR_LIMB, with the
60 			 * most significant bit (with weight 2**N) implicit.
61 			 *
62 			 * Special case for DIVISOR_LIMB == 100...000.
63 			 */
64 			if (!(divisor_limb << 1))
65 				divisor_limb_inverted = ~(mpi_limb_t)0;
66 			else
67 				udiv_qrnnd(divisor_limb_inverted, dummy,
68 						-divisor_limb, 0, divisor_limb);
69 
70 			n1 = dividend_ptr[dividend_size - 1];
71 			r = n1 >> (BITS_PER_MPI_LIMB - normalization_steps);
72 
73 			/* Possible optimization:
74 			 * if (r == 0
75 			 * && divisor_limb > ((n1 << normalization_steps)
76 			 *		       | (dividend_ptr[dividend_size - 2] >> ...)))
77 			 * ...one division less...
78 			 */
79 			for (i = dividend_size - 2; i >= 0; i--) {
80 				n0 = dividend_ptr[i];
81 				UDIV_QRNND_PREINV(dummy, r, r,
82 						((n1 << normalization_steps)
83 						 | (n0 >> (BITS_PER_MPI_LIMB - normalization_steps))),
84 						divisor_limb, divisor_limb_inverted);
85 				n1 = n0;
86 			}
87 			UDIV_QRNND_PREINV(dummy, r, r,
88 					n1 << normalization_steps,
89 					divisor_limb, divisor_limb_inverted);
90 			return r >> normalization_steps;
91 		} else {
92 			mpi_limb_t divisor_limb_inverted;
93 
94 			/* Compute (2**2N - 2**N * DIVISOR_LIMB) / DIVISOR_LIMB.  The
95 			 * result is a (N+1)-bit approximation to 1/DIVISOR_LIMB, with the
96 			 * most significant bit (with weight 2**N) implicit.
97 			 *
98 			 * Special case for DIVISOR_LIMB == 100...000.
99 			 */
100 			if (!(divisor_limb << 1))
101 				divisor_limb_inverted = ~(mpi_limb_t)0;
102 			else
103 				udiv_qrnnd(divisor_limb_inverted, dummy,
104 						-divisor_limb, 0, divisor_limb);
105 
106 			i = dividend_size - 1;
107 			r = dividend_ptr[i];
108 
109 			if (r >= divisor_limb)
110 				r = 0;
111 			else
112 				i--;
113 
114 			for ( ; i >= 0; i--) {
115 				n0 = dividend_ptr[i];
116 				UDIV_QRNND_PREINV(dummy, r, r,
117 						n0, divisor_limb, divisor_limb_inverted);
118 			}
119 			return r;
120 		}
121 	} else {
122 		if (UDIV_NEEDS_NORMALIZATION) {
123 			int normalization_steps;
124 
125 			normalization_steps = count_leading_zeros(divisor_limb);
126 			if (normalization_steps) {
127 				divisor_limb <<= normalization_steps;
128 
129 				n1 = dividend_ptr[dividend_size - 1];
130 				r = n1 >> (BITS_PER_MPI_LIMB - normalization_steps);
131 
132 				/* Possible optimization:
133 				 * if (r == 0
134 				 * && divisor_limb > ((n1 << normalization_steps)
135 				 *		   | (dividend_ptr[dividend_size - 2] >> ...)))
136 				 * ...one division less...
137 				 */
138 				for (i = dividend_size - 2; i >= 0; i--) {
139 					n0 = dividend_ptr[i];
140 					udiv_qrnnd(dummy, r, r,
141 						((n1 << normalization_steps)
142 						 | (n0 >> (BITS_PER_MPI_LIMB - normalization_steps))),
143 						divisor_limb);
144 					n1 = n0;
145 				}
146 				udiv_qrnnd(dummy, r, r,
147 						n1 << normalization_steps,
148 						divisor_limb);
149 				return r >> normalization_steps;
150 			}
151 		}
152 		/* No normalization needed, either because udiv_qrnnd doesn't require
153 		 * it, or because DIVISOR_LIMB is already normalized.
154 		 */
155 		i = dividend_size - 1;
156 		r = dividend_ptr[i];
157 
158 		if (r >= divisor_limb)
159 			r = 0;
160 		else
161 			i--;
162 
163 		for (; i >= 0; i--) {
164 			n0 = dividend_ptr[i];
165 			udiv_qrnnd(dummy, r, r, n0, divisor_limb);
166 		}
167 		return r;
168 	}
169 }
170 
171 /* Divide num (NP/NSIZE) by den (DP/DSIZE) and write
172  * the NSIZE-DSIZE least significant quotient limbs at QP
173  * and the DSIZE long remainder at NP.	If QEXTRA_LIMBS is
174  * non-zero, generate that many fraction bits and append them after the
175  * other quotient limbs.
176  * Return the most significant limb of the quotient, this is always 0 or 1.
177  *
178  * Preconditions:
179  * 0. NSIZE >= DSIZE.
180  * 1. The most significant bit of the divisor must be set.
181  * 2. QP must either not overlap with the input operands at all, or
182  *    QP + DSIZE >= NP must hold true.	(This means that it's
183  *    possible to put the quotient in the high part of NUM, right after the
184  *    remainder in NUM.
185  * 3. NSIZE >= DSIZE, even if QEXTRA_LIMBS is non-zero.
186  */
187 
188 mpi_limb_t
189 mpihelp_divrem(mpi_ptr_t qp, mpi_size_t qextra_limbs,
190 	       mpi_ptr_t np, mpi_size_t nsize, mpi_ptr_t dp, mpi_size_t dsize)
191 {
192 	mpi_limb_t most_significant_q_limb = 0;
193 
194 	switch (dsize) {
195 	case 0:
196 		/* We are asked to divide by zero, so go ahead and do it!  (To make
197 		   the compiler not remove this statement, return the value.)  */
198 		/*
199 		 * existing clients of this function have been modified
200 		 * not to call it with dsize == 0, so this should not happen
201 		 */
202 		return 1 / dsize;
203 
204 	case 1:
205 		{
206 			mpi_size_t i;
207 			mpi_limb_t n1;
208 			mpi_limb_t d;
209 
210 			d = dp[0];
211 			n1 = np[nsize - 1];
212 
213 			if (n1 >= d) {
214 				n1 -= d;
215 				most_significant_q_limb = 1;
216 			}
217 
218 			qp += qextra_limbs;
219 			for (i = nsize - 2; i >= 0; i--)
220 				udiv_qrnnd(qp[i], n1, n1, np[i], d);
221 			qp -= qextra_limbs;
222 
223 			for (i = qextra_limbs - 1; i >= 0; i--)
224 				udiv_qrnnd(qp[i], n1, n1, 0, d);
225 
226 			np[0] = n1;
227 		}
228 		break;
229 
230 	case 2:
231 		{
232 			mpi_size_t i;
233 			mpi_limb_t n1, n0, n2;
234 			mpi_limb_t d1, d0;
235 
236 			np += nsize - 2;
237 			d1 = dp[1];
238 			d0 = dp[0];
239 			n1 = np[1];
240 			n0 = np[0];
241 
242 			if (n1 >= d1 && (n1 > d1 || n0 >= d0)) {
243 				sub_ddmmss(n1, n0, n1, n0, d1, d0);
244 				most_significant_q_limb = 1;
245 			}
246 
247 			for (i = qextra_limbs + nsize - 2 - 1; i >= 0; i--) {
248 				mpi_limb_t q;
249 				mpi_limb_t r;
250 
251 				if (i >= qextra_limbs)
252 					np--;
253 				else
254 					np[0] = 0;
255 
256 				if (n1 == d1) {
257 					/* Q should be either 111..111 or 111..110.  Need special
258 					 * treatment of this rare case as normal division would
259 					 * give overflow.  */
260 					q = ~(mpi_limb_t) 0;
261 
262 					r = n0 + d1;
263 					if (r < d1) {	/* Carry in the addition? */
264 						add_ssaaaa(n1, n0, r - d0,
265 							   np[0], 0, d0);
266 						qp[i] = q;
267 						continue;
268 					}
269 					n1 = d0 - (d0 != 0 ? 1 : 0);
270 					n0 = -d0;
271 				} else {
272 					udiv_qrnnd(q, r, n1, n0, d1);
273 					umul_ppmm(n1, n0, d0, q);
274 				}
275 
276 				n2 = np[0];
277 q_test:
278 				if (n1 > r || (n1 == r && n0 > n2)) {
279 					/* The estimated Q was too large.  */
280 					q--;
281 					sub_ddmmss(n1, n0, n1, n0, 0, d0);
282 					r += d1;
283 					if (r >= d1)	/* If not carry, test Q again.  */
284 						goto q_test;
285 				}
286 
287 				qp[i] = q;
288 				sub_ddmmss(n1, n0, r, n2, n1, n0);
289 			}
290 			np[1] = n1;
291 			np[0] = n0;
292 		}
293 		break;
294 
295 	default:
296 		{
297 			mpi_size_t i;
298 			mpi_limb_t dX, d1, n0;
299 
300 			np += nsize - dsize;
301 			dX = dp[dsize - 1];
302 			d1 = dp[dsize - 2];
303 			n0 = np[dsize - 1];
304 
305 			if (n0 >= dX) {
306 				if (n0 > dX
307 				    || mpihelp_cmp(np, dp, dsize - 1) >= 0) {
308 					mpihelp_sub_n(np, np, dp, dsize);
309 					n0 = np[dsize - 1];
310 					most_significant_q_limb = 1;
311 				}
312 			}
313 
314 			for (i = qextra_limbs + nsize - dsize - 1; i >= 0; i--) {
315 				mpi_limb_t q;
316 				mpi_limb_t n1, n2;
317 				mpi_limb_t cy_limb;
318 
319 				if (i >= qextra_limbs) {
320 					np--;
321 					n2 = np[dsize];
322 				} else {
323 					n2 = np[dsize - 1];
324 					MPN_COPY_DECR(np + 1, np, dsize - 1);
325 					np[0] = 0;
326 				}
327 
328 				if (n0 == dX) {
329 					/* This might over-estimate q, but it's probably not worth
330 					 * the extra code here to find out.  */
331 					q = ~(mpi_limb_t) 0;
332 				} else {
333 					mpi_limb_t r;
334 
335 					udiv_qrnnd(q, r, n0, np[dsize - 1], dX);
336 					umul_ppmm(n1, n0, d1, q);
337 
338 					while (n1 > r
339 					       || (n1 == r
340 						   && n0 > np[dsize - 2])) {
341 						q--;
342 						r += dX;
343 						if (r < dX)	/* I.e. "carry in previous addition?" */
344 							break;
345 						n1 -= n0 < d1;
346 						n0 -= d1;
347 					}
348 				}
349 
350 				/* Possible optimization: We already have (q * n0) and (1 * n1)
351 				 * after the calculation of q.  Taking advantage of that, we
352 				 * could make this loop make two iterations less.  */
353 				cy_limb = mpihelp_submul_1(np, dp, dsize, q);
354 
355 				if (n2 != cy_limb) {
356 					mpihelp_add_n(np, np, dp, dsize);
357 					q--;
358 				}
359 
360 				qp[i] = q;
361 				n0 = np[dsize - 1];
362 			}
363 		}
364 	}
365 
366 	return most_significant_q_limb;
367 }
368 
369 /****************
370  * Divide (DIVIDEND_PTR,,DIVIDEND_SIZE) by DIVISOR_LIMB.
371  * Write DIVIDEND_SIZE limbs of quotient at QUOT_PTR.
372  * Return the single-limb remainder.
373  * There are no constraints on the value of the divisor.
374  *
375  * QUOT_PTR and DIVIDEND_PTR might point to the same limb.
376  */
377 
378 mpi_limb_t
379 mpihelp_divmod_1(mpi_ptr_t quot_ptr,
380 		mpi_ptr_t dividend_ptr, mpi_size_t dividend_size,
381 		mpi_limb_t divisor_limb)
382 {
383 	mpi_size_t i;
384 	mpi_limb_t n1, n0, r;
385 	mpi_limb_t dummy __maybe_unused;
386 
387 	if (!dividend_size)
388 		return 0;
389 
390 	/* If multiplication is much faster than division, and the
391 	 * dividend is large, pre-invert the divisor, and use
392 	 * only multiplications in the inner loop.
393 	 *
394 	 * This test should be read:
395 	 * Does it ever help to use udiv_qrnnd_preinv?
396 	 * && Does what we save compensate for the inversion overhead?
397 	 */
398 	if (UDIV_TIME > (2 * UMUL_TIME + 6)
399 			&& (UDIV_TIME - (2 * UMUL_TIME + 6)) * dividend_size > UDIV_TIME) {
400 		int normalization_steps;
401 
402 		normalization_steps = count_leading_zeros(divisor_limb);
403 		if (normalization_steps) {
404 			mpi_limb_t divisor_limb_inverted;
405 
406 			divisor_limb <<= normalization_steps;
407 
408 			/* Compute (2**2N - 2**N * DIVISOR_LIMB) / DIVISOR_LIMB.  The
409 			 * result is a (N+1)-bit approximation to 1/DIVISOR_LIMB, with the
410 			 * most significant bit (with weight 2**N) implicit.
411 			 */
412 			/* Special case for DIVISOR_LIMB == 100...000.  */
413 			if (!(divisor_limb << 1))
414 				divisor_limb_inverted = ~(mpi_limb_t)0;
415 			else
416 				udiv_qrnnd(divisor_limb_inverted, dummy,
417 						-divisor_limb, 0, divisor_limb);
418 
419 			n1 = dividend_ptr[dividend_size - 1];
420 			r = n1 >> (BITS_PER_MPI_LIMB - normalization_steps);
421 
422 			/* Possible optimization:
423 			 * if (r == 0
424 			 * && divisor_limb > ((n1 << normalization_steps)
425 			 *		       | (dividend_ptr[dividend_size - 2] >> ...)))
426 			 * ...one division less...
427 			 */
428 			for (i = dividend_size - 2; i >= 0; i--) {
429 				n0 = dividend_ptr[i];
430 				UDIV_QRNND_PREINV(quot_ptr[i + 1], r, r,
431 						((n1 << normalization_steps)
432 						 | (n0 >> (BITS_PER_MPI_LIMB - normalization_steps))),
433 						divisor_limb, divisor_limb_inverted);
434 				n1 = n0;
435 			}
436 			UDIV_QRNND_PREINV(quot_ptr[0], r, r,
437 					n1 << normalization_steps,
438 					divisor_limb, divisor_limb_inverted);
439 			return r >> normalization_steps;
440 		} else {
441 			mpi_limb_t divisor_limb_inverted;
442 
443 			/* Compute (2**2N - 2**N * DIVISOR_LIMB) / DIVISOR_LIMB.  The
444 			 * result is a (N+1)-bit approximation to 1/DIVISOR_LIMB, with the
445 			 * most significant bit (with weight 2**N) implicit.
446 			 */
447 			/* Special case for DIVISOR_LIMB == 100...000.  */
448 			if (!(divisor_limb << 1))
449 				divisor_limb_inverted = ~(mpi_limb_t) 0;
450 			else
451 				udiv_qrnnd(divisor_limb_inverted, dummy,
452 						-divisor_limb, 0, divisor_limb);
453 
454 			i = dividend_size - 1;
455 			r = dividend_ptr[i];
456 
457 			if (r >= divisor_limb)
458 				r = 0;
459 			else
460 				quot_ptr[i--] = 0;
461 
462 			for ( ; i >= 0; i--) {
463 				n0 = dividend_ptr[i];
464 				UDIV_QRNND_PREINV(quot_ptr[i], r, r,
465 						n0, divisor_limb, divisor_limb_inverted);
466 			}
467 			return r;
468 		}
469 	} else {
470 		if (UDIV_NEEDS_NORMALIZATION) {
471 			int normalization_steps;
472 
473 			normalization_steps = count_leading_zeros(divisor_limb);
474 			if (normalization_steps) {
475 				divisor_limb <<= normalization_steps;
476 
477 				n1 = dividend_ptr[dividend_size - 1];
478 				r = n1 >> (BITS_PER_MPI_LIMB - normalization_steps);
479 
480 				/* Possible optimization:
481 				 * if (r == 0
482 				 * && divisor_limb > ((n1 << normalization_steps)
483 				 *		   | (dividend_ptr[dividend_size - 2] >> ...)))
484 				 * ...one division less...
485 				 */
486 				for (i = dividend_size - 2; i >= 0; i--) {
487 					n0 = dividend_ptr[i];
488 					udiv_qrnnd(quot_ptr[i + 1], r, r,
489 						((n1 << normalization_steps)
490 						 | (n0 >> (BITS_PER_MPI_LIMB - normalization_steps))),
491 						divisor_limb);
492 					n1 = n0;
493 				}
494 				udiv_qrnnd(quot_ptr[0], r, r,
495 						n1 << normalization_steps,
496 						divisor_limb);
497 				return r >> normalization_steps;
498 			}
499 		}
500 		/* No normalization needed, either because udiv_qrnnd doesn't require
501 		 * it, or because DIVISOR_LIMB is already normalized.
502 		 */
503 		i = dividend_size - 1;
504 		r = dividend_ptr[i];
505 
506 		if (r >= divisor_limb)
507 			r = 0;
508 		else
509 			quot_ptr[i--] = 0;
510 
511 		for (; i >= 0; i--) {
512 			n0 = dividend_ptr[i];
513 			udiv_qrnnd(quot_ptr[i], r, r, n0, divisor_limb);
514 		}
515 		return r;
516 	}
517 }
518