xref: /freebsd/lib/msun/bsdsrc/b_tgamma.c (revision 55141f2c8991b2a6adbf30bb0fe3e6cbc303f06d)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1992, 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 /*
33  * The original code, FreeBSD's old svn r93211, contained the following
34  * attribution:
35  *
36  *    This code by P. McIlroy, Oct 1992;
37  *
38  *    The financial support of UUNET Communications Services is greatfully
39  *    acknowledged.
40  *
41  *  The algorithm remains, but the code has been re-arranged to facilitate
42  *  porting to other precisions.
43  */
44 
45 /* @(#)gamma.c	8.1 (Berkeley) 6/4/93 */
46 #include <sys/cdefs.h>
47 #include <float.h>
48 
49 #include "math.h"
50 #include "math_private.h"
51 
52 /* Used in b_log.c and below. */
53 struct Double {
54 	double a;
55 	double b;
56 };
57 
58 #include "b_log.c"
59 #include "b_exp.c"
60 
61 /*
62  * The range is broken into several subranges.  Each is handled by its
63  * helper functions.
64  *
65  *         x >=   6.0: large_gam(x)
66  *   6.0 > x >= xleft: small_gam(x) where xleft = 1 + left + x0.
67  * xleft > x >   iota: smaller_gam(x) where iota = 1e-17.
68  *  iota > x >  -itoa: Handle x near 0.
69  * -iota > x         : neg_gam
70  *
71  * Special values:
72  *	-Inf:			return NaN and raise invalid;
73  *	negative integer:	return NaN and raise invalid;
74  *	other x ~< 177.79:	return +-0 and raise underflow;
75  *	+-0:			return +-Inf and raise divide-by-zero;
76  *	finite x ~> 171.63:	return +Inf and raise overflow;
77  *	+Inf:			return +Inf;
78  *	NaN: 			return NaN.
79  *
80  * Accuracy: tgamma(x) is accurate to within
81  *	x > 0:  error provably < 0.9ulp.
82  *	Maximum observed in 1,000,000 trials was .87ulp.
83  *	x < 0:
84  *	Maximum observed error < 4ulp in 1,000,000 trials.
85  */
86 
87 /*
88  * Constants for large x approximation (x in [6, Inf])
89  * (Accurate to 2.8*10^-19 absolute)
90  */
91 
92 static const double zero = 0.;
93 static const volatile double tiny = 1e-300;
94 /*
95  * x >= 6
96  *
97  * Use the asymptotic approximation (Stirling's formula) adjusted fof
98  * equal-ripples:
99  *
100  * log(G(x)) ~= (x-0.5)*(log(x)-1) + 0.5(log(2*pi)-1) + 1/x*P(1/(x*x))
101  *
102  * Keep extra precision in multiplying (x-.5)(log(x)-1), to avoid
103  * premature round-off.
104  *
105  * Accurate to max(ulp(1/128) absolute, 2^-66 relative) error.
106  */
107 static const double
108     ln2pi_hi =  0.41894531250000000,
109     ln2pi_lo = -6.7792953272582197e-6,
110     Pa0 =  8.3333333333333329e-02, /* 0x3fb55555, 0x55555555 */
111     Pa1 = -2.7777777777735404e-03, /* 0xbf66c16c, 0x16c145ec */
112     Pa2 =  7.9365079044114095e-04, /* 0x3f4a01a0, 0x183de82d */
113     Pa3 = -5.9523715464225254e-04, /* 0xbf438136, 0x0e681f62 */
114     Pa4 =  8.4161391899445698e-04, /* 0x3f4b93f8, 0x21042a13 */
115     Pa5 = -1.9065246069191080e-03, /* 0xbf5f3c8b, 0x357cb64e */
116     Pa6 =  5.9047708485785158e-03, /* 0x3f782f99, 0xdaf5d65f */
117     Pa7 = -1.6484018705183290e-02; /* 0xbf90e12f, 0xc4fb4df0 */
118 
119 static struct Double
120 large_gam(double x)
121 {
122 	double p, z, thi, tlo, xhi, xlo;
123 	struct Double u;
124 
125 	z = 1 / (x * x);
126 	p = Pa0 + z * (Pa1 + z * (Pa2 + z * (Pa3 + z * (Pa4 + z * (Pa5 +
127 	    z * (Pa6 + z * Pa7))))));
128 	p = p / x;
129 
130 	u = __log__D(x);
131 	u.a -= 1;
132 
133 	/* Split (x - 0.5) in high and low parts. */
134 	x -= 0.5;
135 	xhi = (float)x;
136 	xlo = x - xhi;
137 
138 	/* Compute  t = (x-.5)*(log(x)-1) in extra precision. */
139 	thi = xhi * u.a;
140 	tlo = xlo * u.a + x * u.b;
141 
142 	/* Compute thi + tlo + ln2pi_hi + ln2pi_lo + p. */
143 	tlo += ln2pi_lo;
144 	tlo += p;
145 	u.a = ln2pi_hi + tlo;
146 	u.a += thi;
147 	u.b = thi - u.a;
148 	u.b += ln2pi_hi;
149 	u.b += tlo;
150 	return (u);
151 }
152 /*
153  * Rational approximation, A0 + x * x * P(x) / Q(x), on the interval
154  * [1.066.., 2.066..] accurate to 4.25e-19.
155  *
156  * Returns r.a + r.b = a0 + (z + c)^2 * p / q, with r.a truncated.
157  */
158 static const double
159 #if 0
160     a0_hi =  8.8560319441088875e-1,
161     a0_lo = -4.9964270364690197e-17,
162 #else
163     a0_hi =  8.8560319441088875e-01, /* 0x3fec56dc, 0x82a74aef */
164     a0_lo = -4.9642368725563397e-17, /* 0xbc8c9deb, 0xaa64afc3 */
165 #endif
166     P0 =  6.2138957182182086e-1,
167     P1 =  2.6575719865153347e-1,
168     P2 =  5.5385944642991746e-3,
169     P3 =  1.3845669830409657e-3,
170     P4 =  2.4065995003271137e-3,
171     Q0 =  1.4501953125000000e+0,
172     Q1 =  1.0625852194801617e+0,
173     Q2 = -2.0747456194385994e-1,
174     Q3 = -1.4673413178200542e-1,
175     Q4 =  3.0787817615617552e-2,
176     Q5 =  5.1244934798066622e-3,
177     Q6 = -1.7601274143166700e-3,
178     Q7 =  9.3502102357378894e-5,
179     Q8 =  6.1327550747244396e-6;
180 
181 static struct Double
182 ratfun_gam(double z, double c)
183 {
184 	double p, q, thi, tlo;
185 	struct Double r;
186 
187 	q = Q0 + z * (Q1 + z * (Q2 + z * (Q3 + z * (Q4 + z * (Q5 +
188 	    z * (Q6 + z * (Q7 + z * Q8)))))));
189 	p = P0 + z * (P1 + z * (P2 + z * (P3 + z * P4)));
190 	p = p / q;
191 
192 	/* Split z into high and low parts. */
193 	thi = (float)z;
194 	tlo = (z - thi) + c;
195 	tlo *= (thi + z);
196 
197 	/* Split (z+c)^2 into high and low parts. */
198 	thi *= thi;
199 	q = thi;
200 	thi = (float)thi;
201 	tlo += (q - thi);
202 
203 	/* Split p/q into high and low parts. */
204 	r.a = (float)p;
205 	r.b = p - r.a;
206 
207 	tlo = tlo * p + thi * r.b + a0_lo;
208 	thi *= r.a;				/* t = (z+c)^2*(P/Q) */
209 	r.a = (float)(thi + a0_hi);
210 	r.b = ((a0_hi - r.a) + thi) + tlo;
211 	return (r);				/* r = a0 + t */
212 }
213 /*
214  * x < 6
215  *
216  * Use argument reduction G(x+1) = xG(x) to reach the range [1.066124,
217  * 2.066124].  Use a rational approximation centered at the minimum
218  * (x0+1) to ensure monotonicity.
219  *
220  * Good to < 1 ulp.  (provably .90 ulp; .87 ulp on 1,000,000 runs.)
221  * It also has correct monotonicity.
222  */
223 static const double
224     left = -0.3955078125,	/* left boundary for rat. approx */
225     x0 = 4.6163214496836236e-1;	/* xmin - 1 */
226 
227 static double
228 small_gam(double x)
229 {
230 	double t, y, ym1;
231 	struct Double yy, r;
232 
233 	y = x - 1;
234 	if (y <= 1 + (left + x0)) {
235 		yy = ratfun_gam(y - x0, 0);
236 		return (yy.a + yy.b);
237 	}
238 
239 	r.a = (float)y;
240 	yy.a = r.a - 1;
241 	y = y - 1 ;
242 	r.b = yy.b = y - yy.a;
243 
244 	/* Argument reduction: G(x+1) = x*G(x) */
245 	for (ym1 = y - 1; ym1 > left + x0; y = ym1--, yy.a--) {
246 		t = r.a * yy.a;
247 		r.b = r.a * yy.b + y * r.b;
248 		r.a = (float)t;
249 		r.b += (t - r.a);
250 	}
251 
252 	/* Return r*tgamma(y). */
253 	yy = ratfun_gam(y - x0, 0);
254 	y = r.b * (yy.a + yy.b) + r.a * yy.b;
255 	y += yy.a * r.a;
256 	return (y);
257 }
258 /*
259  * Good on (0, 1+x0+left].  Accurate to 1 ulp.
260  */
261 static double
262 smaller_gam(double x)
263 {
264 	double d, rhi, rlo, t, xhi, xlo;
265 	struct Double r;
266 
267 	if (x < x0 + left) {
268 		t = (float)x;
269 		d = (t + x) * (x - t);
270 		t *= t;
271 		xhi = (float)(t + x);
272 		xlo = x - xhi;
273 		xlo += t;
274 		xlo += d;
275 		t = 1 - x0;
276 		t += x;
277 		d = 1 - x0;
278 		d -= t;
279 		d += x;
280 		x = xhi + xlo;
281 	} else {
282 		xhi = (float)x;
283 		xlo = x - xhi;
284 		t = x - x0;
285 		d = - x0 - t;
286 		d += x;
287 	}
288 
289 	r = ratfun_gam(t, d);
290 	d = (float)(r.a / x);
291 	r.a -= d * xhi;
292 	r.a -= d * xlo;
293 	r.a += r.b;
294 
295 	return (d + r.a / x);
296 }
297 /*
298  * x < 0
299  *
300  * Use reflection formula, G(x) = pi/(sin(pi*x)*x*G(x)).
301  * At negative integers, return NaN and raise invalid.
302  */
303 static double
304 neg_gam(double x)
305 {
306 	int sgn = 1;
307 	struct Double lg, lsine;
308 	double y, z;
309 
310 	y = ceil(x);
311 	if (y == x)		/* Negative integer. */
312 		return ((x - x) / zero);
313 
314 	z = y - x;
315 	if (z > 0.5)
316 		z = 1 - z;
317 
318 	y = y / 2;
319 	if (y == ceil(y))
320 		sgn = -1;
321 
322 	if (z < 0.25)
323 		z = sinpi(z);
324 	else
325 		z = cospi(0.5 - z);
326 
327 	/* Special case: G(1-x) = Inf; G(x) may be nonzero. */
328 	if (x < -170) {
329 
330 		if (x < -190)
331 			return (sgn * tiny * tiny);
332 
333 		y = 1 - x;			/* exact: 128 < |x| < 255 */
334 		lg = large_gam(y);
335 		lsine = __log__D(M_PI / z);	/* = TRUNC(log(u)) + small */
336 		lg.a -= lsine.a;		/* exact (opposite signs) */
337 		lg.b -= lsine.b;
338 		y = -(lg.a + lg.b);
339 		z = (y + lg.a) + lg.b;
340 		y = __exp__D(y, z);
341 		if (sgn < 0) y = -y;
342 		return (y);
343 	}
344 
345 	y = 1 - x;
346 	if (1 - y == x)
347 		y = tgamma(y);
348 	else		/* 1-x is inexact */
349 		y = - x * tgamma(-x);
350 
351 	if (sgn < 0) y = -y;
352 	return (M_PI / (y * z));
353 }
354 /*
355  * xmax comes from lgamma(xmax) - emax * log(2) = 0.
356  * static const float  xmax = 35.040095f
357  * static const double xmax = 171.624376956302725;
358  * ld80: LD80C(0xdb718c066b352e20, 10, 1.75554834290446291689e+03L),
359  * ld128: 1.75554834290446291700388921607020320e+03L,
360  *
361  * iota is a sloppy threshold to isolate x = 0.
362  */
363 static const double xmax = 171.624376956302725;
364 static const double iota = 0x1p-56;
365 
366 double
367 tgamma(double x)
368 {
369 	struct Double u;
370 
371 	if (x >= 6) {
372 		if (x > xmax)
373 			return (x / zero);
374 		u = large_gam(x);
375 		return (__exp__D(u.a, u.b));
376 	}
377 
378 	if (x >= 1 + left + x0)
379 		return (small_gam(x));
380 
381 	if (x > iota)
382 		return (smaller_gam(x));
383 
384 	if (x > -iota) {
385 		if (x != 0.)
386 			u.a = 1 - tiny;	/* raise inexact */
387 		return (1 / x);
388 	}
389 
390 	if (!isfinite(x))
391 		return (x - x);		/* x is NaN or -Inf */
392 
393 	return (neg_gam(x));
394 }
395 
396 #if (LDBL_MANT_DIG == 53)
397 __weak_reference(tgamma, tgammal);
398 #endif
399