xref: /freebsd/lib/msun/src/e_lgamma_r.c (revision b197d4b893974c9eb4d7b38704c6d5c486235d6f)
1 /* @(#)e_lgamma_r.c 1.3 95/01/18 */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunSoft, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 #include <sys/cdefs.h>
14 __FBSDID("$FreeBSD$");
15 
16 /* __ieee754_lgamma_r(x, signgamp)
17  * Reentrant version of the logarithm of the Gamma function
18  * with user provide pointer for the sign of Gamma(x).
19  *
20  * Method:
21  *   1. Argument Reduction for 0 < x <= 8
22  * 	Since gamma(1+s)=s*gamma(s), for x in [0,8], we may
23  * 	reduce x to a number in [1.5,2.5] by
24  * 		lgamma(1+s) = log(s) + lgamma(s)
25  *	for example,
26  *		lgamma(7.3) = log(6.3) + lgamma(6.3)
27  *			    = log(6.3*5.3) + lgamma(5.3)
28  *			    = log(6.3*5.3*4.3*3.3*2.3) + lgamma(2.3)
29  *   2. Polynomial approximation of lgamma around its
30  *	minimum ymin=1.461632144968362245 to maintain monotonicity.
31  *	On [ymin-0.23, ymin+0.27] (i.e., [1.23164,1.73163]), use
32  *		Let z = x-ymin;
33  *		lgamma(x) = -1.214862905358496078218 + z^2*poly(z)
34  *	where
35  *		poly(z) is a 14 degree polynomial.
36  *   2. Rational approximation in the primary interval [2,3]
37  *	We use the following approximation:
38  *		s = x-2.0;
39  *		lgamma(x) = 0.5*s + s*P(s)/Q(s)
40  *	with accuracy
41  *		|P/Q - (lgamma(x)-0.5s)| < 2**-61.71
42  *	Our algorithms are based on the following observation
43  *
44  *                             zeta(2)-1    2    zeta(3)-1    3
45  * lgamma(2+s) = s*(1-Euler) + --------- * s  -  --------- * s  + ...
46  *                                 2                 3
47  *
48  *	where Euler = 0.5771... is the Euler constant, which is very
49  *	close to 0.5.
50  *
51  *   3. For x>=8, we have
52  *	lgamma(x)~(x-0.5)log(x)-x+0.5*log(2pi)+1/(12x)-1/(360x**3)+....
53  *	(better formula:
54  *	   lgamma(x)~(x-0.5)*(log(x)-1)-.5*(log(2pi)-1) + ...)
55  *	Let z = 1/x, then we approximation
56  *		f(z) = lgamma(x) - (x-0.5)(log(x)-1)
57  *	by
58  *	  			    3       5             11
59  *		w = w0 + w1*z + w2*z  + w3*z  + ... + w6*z
60  *	where
61  *		|w - f(z)| < 2**-58.74
62  *
63  *   4. For negative x, since (G is gamma function)
64  *		-x*G(-x)*G(x) = pi/sin(pi*x),
65  * 	we have
66  * 		G(x) = pi/(sin(pi*x)*(-x)*G(-x))
67  *	since G(-x) is positive, sign(G(x)) = sign(sin(pi*x)) for x<0
68  *	Hence, for x<0, signgam = sign(sin(pi*x)) and
69  *		lgamma(x) = log(|Gamma(x)|)
70  *			  = log(pi/(|x*sin(pi*x)|)) - lgamma(-x);
71  *	Note: one should avoid compute pi*(-x) directly in the
72  *	      computation of sin(pi*(-x)).
73  *
74  *   5. Special Cases
75  *		lgamma(2+s) ~ s*(1-Euler) for tiny s
76  *		lgamma(1) = lgamma(2) = 0
77  *		lgamma(x) ~ -log(|x|) for tiny x
78  *		lgamma(0) = lgamma(neg.integer) = inf and raise divide-by-zero
79  *		lgamma(inf) = inf
80  *		lgamma(-inf) = inf (bug for bug compatible with C99!?)
81  */
82 
83 #include <float.h>
84 
85 #include "math.h"
86 #include "math_private.h"
87 
88 static const volatile double vzero = 0;
89 
90 static const double
91 zero=  0.00000000000000000000e+00,
92 half=  5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */
93 one =  1.00000000000000000000e+00, /* 0x3FF00000, 0x00000000 */
94 pi  =  3.14159265358979311600e+00, /* 0x400921FB, 0x54442D18 */
95 a0  =  7.72156649015328655494e-02, /* 0x3FB3C467, 0xE37DB0C8 */
96 a1  =  3.22467033424113591611e-01, /* 0x3FD4A34C, 0xC4A60FAD */
97 a2  =  6.73523010531292681824e-02, /* 0x3FB13E00, 0x1A5562A7 */
98 a3  =  2.05808084325167332806e-02, /* 0x3F951322, 0xAC92547B */
99 a4  =  7.38555086081402883957e-03, /* 0x3F7E404F, 0xB68FEFE8 */
100 a5  =  2.89051383673415629091e-03, /* 0x3F67ADD8, 0xCCB7926B */
101 a6  =  1.19270763183362067845e-03, /* 0x3F538A94, 0x116F3F5D */
102 a7  =  5.10069792153511336608e-04, /* 0x3F40B6C6, 0x89B99C00 */
103 a8  =  2.20862790713908385557e-04, /* 0x3F2CF2EC, 0xED10E54D */
104 a9  =  1.08011567247583939954e-04, /* 0x3F1C5088, 0x987DFB07 */
105 a10 =  2.52144565451257326939e-05, /* 0x3EFA7074, 0x428CFA52 */
106 a11 =  4.48640949618915160150e-05, /* 0x3F07858E, 0x90A45837 */
107 tc  =  1.46163214496836224576e+00, /* 0x3FF762D8, 0x6356BE3F */
108 tf  = -1.21486290535849611461e-01, /* 0xBFBF19B9, 0xBCC38A42 */
109 /* tt = -(tail of tf) */
110 tt  = -3.63867699703950536541e-18, /* 0xBC50C7CA, 0xA48A971F */
111 t0  =  4.83836122723810047042e-01, /* 0x3FDEF72B, 0xC8EE38A2 */
112 t1  = -1.47587722994593911752e-01, /* 0xBFC2E427, 0x8DC6C509 */
113 t2  =  6.46249402391333854778e-02, /* 0x3FB08B42, 0x94D5419B */
114 t3  = -3.27885410759859649565e-02, /* 0xBFA0C9A8, 0xDF35B713 */
115 t4  =  1.79706750811820387126e-02, /* 0x3F9266E7, 0x970AF9EC */
116 t5  = -1.03142241298341437450e-02, /* 0xBF851F9F, 0xBA91EC6A */
117 t6  =  6.10053870246291332635e-03, /* 0x3F78FCE0, 0xE370E344 */
118 t7  = -3.68452016781138256760e-03, /* 0xBF6E2EFF, 0xB3E914D7 */
119 t8  =  2.25964780900612472250e-03, /* 0x3F6282D3, 0x2E15C915 */
120 t9  = -1.40346469989232843813e-03, /* 0xBF56FE8E, 0xBF2D1AF1 */
121 t10 =  8.81081882437654011382e-04, /* 0x3F4CDF0C, 0xEF61A8E9 */
122 t11 = -5.38595305356740546715e-04, /* 0xBF41A610, 0x9C73E0EC */
123 t12 =  3.15632070903625950361e-04, /* 0x3F34AF6D, 0x6C0EBBF7 */
124 t13 = -3.12754168375120860518e-04, /* 0xBF347F24, 0xECC38C38 */
125 t14 =  3.35529192635519073543e-04, /* 0x3F35FD3E, 0xE8C2D3F4 */
126 u0  = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */
127 u1  =  6.32827064025093366517e-01, /* 0x3FE4401E, 0x8B005DFF */
128 u2  =  1.45492250137234768737e+00, /* 0x3FF7475C, 0xD119BD6F */
129 u3  =  9.77717527963372745603e-01, /* 0x3FEF4976, 0x44EA8450 */
130 u4  =  2.28963728064692451092e-01, /* 0x3FCD4EAE, 0xF6010924 */
131 u5  =  1.33810918536787660377e-02, /* 0x3F8B678B, 0xBF2BAB09 */
132 v1  =  2.45597793713041134822e+00, /* 0x4003A5D7, 0xC2BD619C */
133 v2  =  2.12848976379893395361e+00, /* 0x40010725, 0xA42B18F5 */
134 v3  =  7.69285150456672783825e-01, /* 0x3FE89DFB, 0xE45050AF */
135 v4  =  1.04222645593369134254e-01, /* 0x3FBAAE55, 0xD6537C88 */
136 v5  =  3.21709242282423911810e-03, /* 0x3F6A5ABB, 0x57D0CF61 */
137 s0  = -7.72156649015328655494e-02, /* 0xBFB3C467, 0xE37DB0C8 */
138 s1  =  2.14982415960608852501e-01, /* 0x3FCB848B, 0x36E20878 */
139 s2  =  3.25778796408930981787e-01, /* 0x3FD4D98F, 0x4F139F59 */
140 s3  =  1.46350472652464452805e-01, /* 0x3FC2BB9C, 0xBEE5F2F7 */
141 s4  =  2.66422703033638609560e-02, /* 0x3F9B481C, 0x7E939961 */
142 s5  =  1.84028451407337715652e-03, /* 0x3F5E26B6, 0x7368F239 */
143 s6  =  3.19475326584100867617e-05, /* 0x3F00BFEC, 0xDD17E945 */
144 r1  =  1.39200533467621045958e+00, /* 0x3FF645A7, 0x62C4AB74 */
145 r2  =  7.21935547567138069525e-01, /* 0x3FE71A18, 0x93D3DCDC */
146 r3  =  1.71933865632803078993e-01, /* 0x3FC601ED, 0xCCFBDF27 */
147 r4  =  1.86459191715652901344e-02, /* 0x3F9317EA, 0x742ED475 */
148 r5  =  7.77942496381893596434e-04, /* 0x3F497DDA, 0xCA41A95B */
149 r6  =  7.32668430744625636189e-06, /* 0x3EDEBAF7, 0xA5B38140 */
150 w0  =  4.18938533204672725052e-01, /* 0x3FDACFE3, 0x90C97D69 */
151 w1  =  8.33333333333329678849e-02, /* 0x3FB55555, 0x5555553B */
152 w2  = -2.77777777728775536470e-03, /* 0xBF66C16C, 0x16B02E5C */
153 w3  =  7.93650558643019558500e-04, /* 0x3F4A019F, 0x98CF38B6 */
154 w4  = -5.95187557450339963135e-04, /* 0xBF4380CB, 0x8C0FE741 */
155 w5  =  8.36339918996282139126e-04, /* 0x3F4B67BA, 0x4CDAD5D1 */
156 w6  = -1.63092934096575273989e-03; /* 0xBF5AB89D, 0x0B9E43E4 */
157 
158 /*
159  * Compute sin(pi*x) without actually doing the pi*x multiplication.
160  * sin_pi(x) is only called for x < 0 and |x| < 2**(p-1) where p is
161  * the precision of x.
162  */
163 static double
164 sin_pi(double x)
165 {
166 	volatile double vz;
167 	double y,z;
168 	int n;
169 
170 	y = -x;
171 
172 	vz = y+0x1p52;			/* depend on 0 <= y < 0x1p52 */
173 	z = vz-0x1p52;			/* rint(y) for the above range */
174 	if (z == y)
175 	    return zero;
176 
177 	vz = y+0x1p50;
178 	GET_LOW_WORD(n,vz);		/* bits for rounded y (units 0.25) */
179 	z = vz-0x1p50;			/* y rounded to a multiple of 0.25 */
180 	if (z > y) {
181 	    z -= 0.25;			/* adjust to round down */
182 	    n--;
183 	}
184 	n &= 7;				/* octant of y mod 2 */
185 	y = y - z + n * 0.25;		/* y mod 2 */
186 
187 	switch (n) {
188 	    case 0:   y =  __kernel_sin(pi*y,zero,0); break;
189 	    case 1:
190 	    case 2:   y =  __kernel_cos(pi*(0.5-y),zero); break;
191 	    case 3:
192 	    case 4:   y =  __kernel_sin(pi*(one-y),zero,0); break;
193 	    case 5:
194 	    case 6:   y = -__kernel_cos(pi*(y-1.5),zero); break;
195 	    default:  y =  __kernel_sin(pi*(y-2.0),zero,0); break;
196 	    }
197 	return -y;
198 }
199 
200 
201 double
202 __ieee754_lgamma_r(double x, int *signgamp)
203 {
204 	double nadj,p,p1,p2,p3,q,r,t,w,y,z;
205 	int32_t hx;
206 	int i,ix,lx;
207 
208 	EXTRACT_WORDS(hx,lx,x);
209 
210     /* purge +-Inf and NaNs */
211 	*signgamp = 1;
212 	ix = hx&0x7fffffff;
213 	if(ix>=0x7ff00000) return x*x;
214 
215     /* purge +-0 and tiny arguments */
216 	*signgamp = 1-2*((uint32_t)hx>>31);
217 	if(ix<0x3c700000) {	/* |x|<2**-56, return -log(|x|) */
218 	    if((ix|lx)==0)
219 	        return one/vzero;
220 	    return -__ieee754_log(fabs(x));
221 	}
222 
223     /* purge negative integers and start evaluation for other x < 0 */
224 	if(hx<0) {
225 	    *signgamp = 1;
226 	    if(ix>=0x43300000) 	/* |x|>=2**52, must be -integer */
227 		return one/vzero;
228 	    t = sin_pi(x);
229 	    if(t==zero) return one/vzero; /* -integer */
230 	    nadj = __ieee754_log(pi/fabs(t*x));
231 	    if(t<zero) *signgamp = -1;
232 	    x = -x;
233 	}
234 
235     /* purge 1 and 2 */
236 	if((((ix-0x3ff00000)|lx)==0)||(((ix-0x40000000)|lx)==0)) r = 0;
237     /* for x < 2.0 */
238 	else if(ix<0x40000000) {
239 	    if(ix<=0x3feccccc) { 	/* lgamma(x) = lgamma(x+1)-log(x) */
240 		r = -__ieee754_log(x);
241 		if(ix>=0x3FE76944) {y = one-x; i= 0;}
242 		else if(ix>=0x3FCDA661) {y= x-(tc-one); i=1;}
243 	  	else {y = x; i=2;}
244 	    } else {
245 	  	r = zero;
246 	        if(ix>=0x3FFBB4C3) {y=2.0-x;i=0;} /* [1.7316,2] */
247 	        else if(ix>=0x3FF3B4C4) {y=x-tc;i=1;} /* [1.23,1.73] */
248 		else {y=x-one;i=2;}
249 	    }
250 	    switch(i) {
251 	      case 0:
252 		z = y*y;
253 		p1 = a0+z*(a2+z*(a4+z*(a6+z*(a8+z*a10))));
254 		p2 = z*(a1+z*(a3+z*(a5+z*(a7+z*(a9+z*a11)))));
255 		p  = y*p1+p2;
256 		r  += p-y/2; break;
257 	      case 1:
258 		z = y*y;
259 		w = z*y;
260 		p1 = t0+w*(t3+w*(t6+w*(t9 +w*t12)));	/* parallel comp */
261 		p2 = t1+w*(t4+w*(t7+w*(t10+w*t13)));
262 		p3 = t2+w*(t5+w*(t8+w*(t11+w*t14)));
263 		p  = z*p1-(tt-w*(p2+y*p3));
264 		r += tf + p; break;
265 	      case 2:
266 		p1 = y*(u0+y*(u1+y*(u2+y*(u3+y*(u4+y*u5)))));
267 		p2 = one+y*(v1+y*(v2+y*(v3+y*(v4+y*v5))));
268 		r += p1/p2-y/2;
269 	    }
270 	}
271     /* x < 8.0 */
272 	else if(ix<0x40200000) {
273 	    i = x;
274 	    y = x-i;
275 	    p = y*(s0+y*(s1+y*(s2+y*(s3+y*(s4+y*(s5+y*s6))))));
276 	    q = one+y*(r1+y*(r2+y*(r3+y*(r4+y*(r5+y*r6)))));
277 	    r = y/2+p/q;
278 	    z = one;	/* lgamma(1+s) = log(s) + lgamma(s) */
279 	    switch(i) {
280 	    case 7: z *= (y+6);		/* FALLTHRU */
281 	    case 6: z *= (y+5);		/* FALLTHRU */
282 	    case 5: z *= (y+4);		/* FALLTHRU */
283 	    case 4: z *= (y+3);		/* FALLTHRU */
284 	    case 3: z *= (y+2);		/* FALLTHRU */
285 		    r += __ieee754_log(z); break;
286 	    }
287     /* 8.0 <= x < 2**56 */
288 	} else if (ix < 0x43700000) {
289 	    t = __ieee754_log(x);
290 	    z = one/x;
291 	    y = z*z;
292 	    w = w0+z*(w1+y*(w2+y*(w3+y*(w4+y*(w5+y*w6)))));
293 	    r = (x-half)*(t-one)+w;
294 	} else
295     /* 2**56 <= x <= inf */
296 	    r =  x*(__ieee754_log(x)-one);
297 	if(hx<0) r = nadj - r;
298 	return r;
299 }
300 
301 #if (LDBL_MANT_DIG == 53)
302 __weak_reference(lgamma_r, lgammal_r);
303 #endif
304