xref: /haiku/src/system/libroot/posix/stdlib/random.c (revision 445d4fd926c569e7b9ae28017da86280aaecbae2)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1983, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <stdint.h>
33 #include <stdlib.h>
34 
35 /*
36  * random.c:
37  *
38  * An improved random number generation package.  In addition to the standard
39  * rand()/srand() like interface, this package also has a special state info
40  * interface.  The initstate() routine is called with a seed, an array of
41  * bytes, and a count of how many bytes are being passed in; this array is
42  * then initialized to contain information for random number generation with
43  * that much state information.  Good sizes for the amount of state
44  * information are 32, 64, 128, and 256 bytes.  The state can be switched by
45  * calling the setstate() routine with the same array as was initiallized
46  * with initstate().  By default, the package runs with 128 bytes of state
47  * information and generates far better random numbers than a linear
48  * congruential generator.  If the amount of state information is less than
49  * 32 bytes, a simple linear congruential R.N.G. is used.
50  *
51  * Internally, the state information is treated as an array of uint32_t's; the
52  * zeroeth element of the array is the type of R.N.G. being used (small
53  * integer); the remainder of the array is the state information for the
54  * R.N.G.  Thus, 32 bytes of state information will give 7 ints worth of
55  * state information, which will allow a degree seven polynomial.  (Note:
56  * the zeroeth word of state information also has some other information
57  * stored in it -- see setstate() for details).
58  *
59  * The random number generation technique is a linear feedback shift register
60  * approach, employing trinomials (since there are fewer terms to sum up that
61  * way).  In this approach, the least significant bit of all the numbers in
62  * the state table will act as a linear feedback shift register, and will
63  * have period 2^deg - 1 (where deg is the degree of the polynomial being
64  * used, assuming that the polynomial is irreducible and primitive).  The
65  * higher order bits will have longer periods, since their values are also
66  * influenced by pseudo-random carries out of the lower bits.  The total
67  * period of the generator is approximately deg*(2**deg - 1); thus doubling
68  * the amount of state information has a vast influence on the period of the
69  * generator.  Note: the deg*(2**deg - 1) is an approximation only good for
70  * large deg, when the period of the shift is the dominant factor.
71  * With deg equal to seven, the period is actually much longer than the
72  * 7*(2**7 - 1) predicted by this formula.
73  *
74  * Modified 28 December 1994 by Jacob S. Rosenberg.
75  * The following changes have been made:
76  * All references to the type u_int have been changed to unsigned long.
77  * All references to type int have been changed to type long.  Other
78  * cleanups have been made as well.  A warning for both initstate and
79  * setstate has been inserted to the effect that on Sparc platforms
80  * the 'arg_state' variable must be forced to begin on word boundaries.
81  * This can be easily done by casting a long integer array to char *.
82  * The overall logic has been left STRICTLY alone.  This software was
83  * tested on both a VAX and Sun SpacsStation with exactly the same
84  * results.  The new version and the original give IDENTICAL results.
85  * The new version is somewhat faster than the original.  As the
86  * documentation says:  "By default, the package runs with 128 bytes of
87  * state information and generates far better random numbers than a linear
88  * congruential generator.  If the amount of state information is less than
89  * 32 bytes, a simple linear congruential R.N.G. is used."  For a buffer of
90  * 128 bytes, this new version runs about 19 percent faster and for a 16
91  * byte buffer it is about 5 percent faster.
92  */
93 
94 /*
95  * For each of the currently supported random number generators, we have a
96  * break value on the amount of state information (you need at least this
97  * many bytes of state info to support this random number generator), a degree
98  * for the polynomial (actually a trinomial) that the R.N.G. is based on, and
99  * the separation between the two lower order coefficients of the trinomial.
100  */
101 #define	TYPE_0		0		/* linear congruential */
102 #define	BREAK_0		8
103 #define	DEG_0		0
104 #define	SEP_0		0
105 
106 #define	TYPE_1		1		/* x**7 + x**3 + 1 */
107 #define	BREAK_1		32
108 #define	DEG_1		7
109 #define	SEP_1		3
110 
111 #define	TYPE_2		2		/* x**15 + x + 1 */
112 #define	BREAK_2		64
113 #define	DEG_2		15
114 #define	SEP_2		1
115 
116 #define	TYPE_3		3		/* x**31 + x**3 + 1 */
117 #define	BREAK_3		128
118 #define	DEG_3		31
119 #define	SEP_3		3
120 
121 #define	TYPE_4		4		/* x**63 + x + 1 */
122 #define	BREAK_4		256
123 #define	DEG_4		63
124 #define	SEP_4		1
125 
126 /*
127  * Array versions of the above information to make code run faster --
128  * relies on fact that TYPE_i == i.
129  */
130 #define	MAX_TYPES	5		/* max number of types above */
131 
132 #define NSHUFF 50       /* to drop some "seed -> 1st value" linearity */
133 
134 static const int degrees[MAX_TYPES] =	{ DEG_0, DEG_1, DEG_2, DEG_3, DEG_4 };
135 static const int seps [MAX_TYPES] =	{ SEP_0, SEP_1, SEP_2, SEP_3, SEP_4 };
136 
137 /*
138  * Initially, everything is set up as if from:
139  *
140  *	initstate(1, randtbl, 128);
141  *
142  * Note that this initialization takes advantage of the fact that srandom()
143  * advances the front and rear pointers 10*rand_deg times, and hence the
144  * rear pointer which starts at 0 will also end up at zero; thus the zeroeth
145  * element of the state information, which contains info about the current
146  * position of the rear pointer is just
147  *
148  *	MAX_TYPES * (rptr - state) + TYPE_3 == TYPE_3.
149  */
150 
151 static uint32_t randtbl[DEG_3 + 1] = {
152 	TYPE_3,
153 	0x2cf41758, 0x27bb3711, 0x4916d4d1, 0x7b02f59f, 0x9b8e28eb, 0xc0e80269,
154 	0x696f5c16, 0x878f1ff5, 0x52d9c07f, 0x916a06cd, 0xb50b3a20, 0x2776970a,
155 	0xee4eb2a6, 0xe94640ec, 0xb1d65612, 0x9d1ed968, 0x1043f6b7, 0xa3432a76,
156 	0x17eacbb9, 0x3c09e2eb, 0x4f8c2b3,  0x708a1f57, 0xee341814, 0x95d0e4d2,
157 	0xb06f216c, 0x8bd2e72e, 0x8f7c38d7, 0xcfc6a8fc, 0x2a59495,  0xa20d2a69,
158 	0xe29d12d1
159 };
160 
161 /*
162  * fptr and rptr are two pointers into the state info, a front and a rear
163  * pointer.  These two pointers are always rand_sep places aparts, as they
164  * cycle cyclically through the state information.  (Yes, this does mean we
165  * could get away with just one pointer, but the code for random() is more
166  * efficient this way).  The pointers are left positioned as they would be
167  * from the call
168  *
169  *	initstate(1, randtbl, 128);
170  *
171  * (The position of the rear pointer, rptr, is really 0 (as explained above
172  * in the initialization of randtbl) because the state table pointer is set
173  * to point to randtbl[1] (as explained below).
174  */
175 static uint32_t *fptr = &randtbl[SEP_3 + 1];
176 static uint32_t *rptr = &randtbl[1];
177 
178 /*
179  * The following things are the pointer to the state information table, the
180  * type of the current generator, the degree of the current polynomial being
181  * used, and the separation between the two pointers.  Note that for efficiency
182  * of random(), we remember the first location of the state information, not
183  * the zeroeth.  Hence it is valid to access state[-1], which is used to
184  * store the type of the R.N.G.  Also, we remember the last location, since
185  * this is more efficient than indexing every time to find the address of
186  * the last element to see if the front and rear pointers have wrapped.
187  */
188 static uint32_t *state = &randtbl[1];
189 static int rand_type = TYPE_3;
190 static int rand_deg = DEG_3;
191 static int rand_sep = SEP_3;
192 static uint32_t *end_ptr = &randtbl[DEG_3 + 1];
193 
194 static inline uint32_t
195 good_rand(uint32_t ctx)
196 {
197 /*
198  * Compute x = (7^5 * x) mod (2^31 - 1)
199  * wihout overflowing 31 bits:
200  *      (2^31 - 1) = 127773 * (7^5) + 2836
201  * From "Random number generators: good ones are hard to find",
202  * Park and Miller, Communications of the ACM, vol. 31, no. 10,
203  * October 1988, p. 1195.
204  */
205 	int32_t hi, lo, x;
206 
207 	/* Transform to [1, 0x7ffffffe] range. */
208 	x = (ctx % 0x7ffffffe) + 1;
209 	hi = x / 127773;
210 	lo = x % 127773;
211 	x = 16807 * lo - 2836 * hi;
212 	if (x < 0)
213 		x += 0x7fffffff;
214 	/* Transform to [0, 0x7ffffffd] range. */
215 	return (x - 1);
216 }
217 
218 /*
219  * srandom:
220  *
221  * Initialize the random number generator based on the given seed.  If the
222  * type is the trivial no-state-information type, just remember the seed.
223  * Otherwise, initializes state[] based on the given "seed" via a linear
224  * congruential generator.  Then, the pointers are set to known locations
225  * that are exactly rand_sep places apart.  Lastly, it cycles the state
226  * information a given number of times to get rid of any initial dependencies
227  * introduced by the L.C.R.N.G.  Note that the initialization of randtbl[]
228  * for default usage relies on values produced by this routine.
229  */
230 void
231 srandom(unsigned int x)
232 {
233 	int i, lim;
234 
235 	state[0] = (uint32_t)x;
236 	if (rand_type == TYPE_0)
237 		lim = NSHUFF;
238 	else {
239 		for (i = 1; i < rand_deg; i++)
240 			state[i] = good_rand(state[i - 1]);
241 		fptr = &state[rand_sep];
242 		rptr = &state[0];
243 		lim = 10 * rand_deg;
244 	}
245 	for (i = 0; i < lim; i++)
246 		(void)random();
247 }
248 
249 /*
250  * srandomdev:
251  *
252  * Many programs choose the seed value in a totally predictable manner.
253  * This often causes problems.  We seed the generator using pseudo-random
254  * data from the kernel.
255  *
256  * Note that this particular seeding procedure can generate states
257  * which are impossible to reproduce by calling srandom() with any
258  * value, since the succeeding terms in the state buffer are no longer
259  * derived from the LC algorithm applied to a fixed seed.
260  */
261 #if 0
262 void
263 srandomdev(void)
264 {
265 	int mib[2];
266 	size_t expected, len;
267 
268 	if (rand_type == TYPE_0)
269 		expected = len = sizeof(state[0]);
270 	else
271 		expected = len = rand_deg * sizeof(state[0]);
272 
273 	mib[0] = CTL_KERN;
274 	mib[1] = KERN_ARND;
275 	if (sysctl(mib, 2, state, &len, NULL, 0) == -1 || len != expected) {
276 		/*
277 		 * The sysctl cannot fail. If it does fail on some FreeBSD
278 		 * derivative or after some future change, just abort so that
279 		 * the problem will be found and fixed. abort is not normally
280 		 * suitable for a library but makes sense here.
281 		 */
282 		abort();
283 	}
284 
285 	if (rand_type != TYPE_0) {
286 		fptr = &state[rand_sep];
287 		rptr = &state[0];
288 	}
289 }
290 #endif
291 
292 /*
293  * initstate:
294  *
295  * Initialize the state information in the given array of n bytes for future
296  * random number generation.  Based on the number of bytes we are given, and
297  * the break values for the different R.N.G.'s, we choose the best (largest)
298  * one we can and set things up for it.  srandom() is then called to
299  * initialize the state information.
300  *
301  * Note that on return from srandom(), we set state[-1] to be the type
302  * multiplexed with the current value of the rear pointer; this is so
303  * successive calls to initstate() won't lose this information and will be
304  * able to restart with setstate().
305  *
306  * Note: the first thing we do is save the current state, if any, just like
307  * setstate() so that it doesn't matter when initstate is called.
308  *
309  * Returns a pointer to the old state.
310  *
311  * Note: The Sparc platform requires that arg_state begin on an int
312  * word boundary; otherwise a bus error will occur. Even so, lint will
313  * complain about mis-alignment, but you should disregard these messages.
314  */
315 char *
316 initstate(unsigned int seed, char *arg_state, size_t n)
317 {
318 	char *ostate = (char *)(&state[-1]);
319 	uint32_t *int_arg_state = (uint32_t *)arg_state;
320 
321 	if (n < BREAK_0)
322 		return (NULL);
323 	if (rand_type == TYPE_0)
324 		state[-1] = rand_type;
325 	else
326 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
327 	if (n < BREAK_1) {
328 		rand_type = TYPE_0;
329 		rand_deg = DEG_0;
330 		rand_sep = SEP_0;
331 	} else if (n < BREAK_2) {
332 		rand_type = TYPE_1;
333 		rand_deg = DEG_1;
334 		rand_sep = SEP_1;
335 	} else if (n < BREAK_3) {
336 		rand_type = TYPE_2;
337 		rand_deg = DEG_2;
338 		rand_sep = SEP_2;
339 	} else if (n < BREAK_4) {
340 		rand_type = TYPE_3;
341 		rand_deg = DEG_3;
342 		rand_sep = SEP_3;
343 	} else {
344 		rand_type = TYPE_4;
345 		rand_deg = DEG_4;
346 		rand_sep = SEP_4;
347 	}
348 	state = int_arg_state + 1; /* first location */
349 	end_ptr = &state[rand_deg];	/* must set end_ptr before srandom */
350 	srandom(seed);
351 	if (rand_type == TYPE_0)
352 		int_arg_state[0] = rand_type;
353 	else
354 		int_arg_state[0] = MAX_TYPES * (rptr - state) + rand_type;
355 	return (ostate);
356 }
357 
358 /*
359  * setstate:
360  *
361  * Restore the state from the given state array.
362  *
363  * Note: it is important that we also remember the locations of the pointers
364  * in the current state information, and restore the locations of the pointers
365  * from the old state information.  This is done by multiplexing the pointer
366  * location into the zeroeth word of the state information.
367  *
368  * Note that due to the order in which things are done, it is OK to call
369  * setstate() with the same state as the current state.
370  *
371  * Returns a pointer to the old state information.
372  *
373  * Note: The Sparc platform requires that arg_state begin on an int
374  * word boundary; otherwise a bus error will occur. Even so, lint will
375  * complain about mis-alignment, but you should disregard these messages.
376  */
377 char *
378 setstate(char *arg_state)
379 {
380 	uint32_t *new_state = (uint32_t *)arg_state;
381 	uint32_t type = new_state[0] % MAX_TYPES;
382 	uint32_t rear = new_state[0] / MAX_TYPES;
383 	char *ostate = (char *)(&state[-1]);
384 
385 	if (type != TYPE_0 && rear >= (uint32_t)degrees[type])
386 		return (NULL);
387 	if (rand_type == TYPE_0)
388 		state[-1] = rand_type;
389 	else
390 		state[-1] = MAX_TYPES * (rptr - state) + rand_type;
391 	rand_type = type;
392 	rand_deg = degrees[type];
393 	rand_sep = seps[type];
394 	state = new_state + 1;
395 	if (rand_type != TYPE_0) {
396 		rptr = &state[rear];
397 		fptr = &state[(rear + rand_sep) % rand_deg];
398 	}
399 	end_ptr = &state[rand_deg];		/* set end_ptr too */
400 	return (ostate);
401 }
402 
403 /*
404  * random:
405  *
406  * If we are using the trivial TYPE_0 R.N.G., just do the old linear
407  * congruential bit.  Otherwise, we do our fancy trinomial stuff, which is
408  * the same in all the other cases due to all the global variables that have
409  * been set up.  The basic operation is to add the number at the rear pointer
410  * into the one at the front pointer.  Then both pointers are advanced to
411  * the next location cyclically in the table.  The value returned is the sum
412  * generated, reduced to 31 bits by throwing away the "least random" low bit.
413  *
414  * Note: the code takes advantage of the fact that both the front and
415  * rear pointers can't wrap on the same call by not testing the rear
416  * pointer if the front one has wrapped.
417  *
418  * Returns a 31-bit random number.
419  */
420 int
421 random(void)
422 {
423 	uint32_t i;
424 	uint32_t *f, *r;
425 
426 	if (rand_type == TYPE_0) {
427 		i = state[0];
428 		state[0] = i = good_rand(i);
429 	} else {
430 		/*
431 		 * Use local variables rather than static variables for speed.
432 		 */
433 		f = fptr; r = rptr;
434 		*f += *r;
435 		i = *f >> 1;	/* chucking least random bit */
436 		if (++f >= end_ptr) {
437 			f = state;
438 			++r;
439 		}
440 		else if (++r >= end_ptr) {
441 			r = state;
442 		}
443 
444 		fptr = f; rptr = r;
445 	}
446 	return ((long)i);
447 }
448