xref: /freebsd/lib/msun/src/s_tanhl.c (revision 02e9120893770924227138ba49df1edb3896112a)
1 /* from: FreeBSD: head/lib/msun/src/s_tanhl.c XXX */
2 
3 /*
4  * ====================================================
5  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6  *
7  * Developed at SunPro, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  */
13 
14 #include <sys/cdefs.h>
15 /*
16  * See s_tanh.c for complete comments.
17  *
18  * Converted to long double by Bruce D. Evans.
19  */
20 
21 #include <float.h>
22 #ifdef __i386__
23 #include <ieeefp.h>
24 #endif
25 
26 #include "math.h"
27 #include "math_private.h"
28 #include "fpmath.h"
29 #include "k_expl.h"
30 
31 #if LDBL_MAX_EXP != 0x4000
32 /* We also require the usual expsign encoding. */
33 #error "Unsupported long double format"
34 #endif
35 
36 #define	BIAS	(LDBL_MAX_EXP - 1)
37 
38 static const volatile double tiny = 1.0e-300;
39 static const double one = 1.0;
40 #if LDBL_MANT_DIG == 64
41 /*
42  * Domain [-0.25, 0.25], range ~[-1.6304e-22, 1.6304e-22]:
43  * |tanh(x)/x - t(x)| < 2**-72.3
44  */
45 static const union IEEEl2bits
46 T3u = LD80C(0xaaaaaaaaaaaaaa9f, -2, -3.33333333333333333017e-1L);
47 #define	T3	T3u.e
48 static const double
49 T5  =  1.3333333333333314e-1,		/*  0x1111111111110a.0p-55 */
50 T7  = -5.3968253968210485e-2,		/* -0x1ba1ba1ba1a1a1.0p-57 */
51 T9  =  2.1869488531393817e-2,		/*  0x1664f488172022.0p-58 */
52 T11 = -8.8632352345964591e-3,		/* -0x1226e34bc138d5.0p-59 */
53 T13 =  3.5921169709993771e-3,		/*  0x1d6d371d3e400f.0p-61 */
54 T15 = -1.4555786415756001e-3,		/* -0x17d923aa63814d.0p-62 */
55 T17 =  5.8645267876296793e-4,		/*  0x13378589b85aa7.0p-63 */
56 T19 = -2.1121033571392224e-4;		/* -0x1baf0af80c4090.0p-65 */
57 #elif LDBL_MANT_DIG == 113
58 /*
59  * Domain [-0.25, 0.25], range ~[-2.4211e-37, 2.4211e-37]:
60  * |tanh(x)/x - t(x)| < 2**121.6
61  */
62 static const long double
63 T3 = -3.33333333333333333333333333333332980e-1L,	/* -0x1555555555555555555555555554e.0p-114L */
64 T5  =  1.33333333333333333333333333332707260e-1L,	/*  0x1111111111111111111111110ab7b.0p-115L */
65 T7  = -5.39682539682539682539682535723482314e-2L,	/* -0x1ba1ba1ba1ba1ba1ba1ba17b5fc98.0p-117L */
66 T9  =  2.18694885361552028218693591149061717e-2L,	/*  0x1664f4882c10f9f32d6b1a12a25e5.0p-118L */
67 T11 = -8.86323552990219656883762347736381851e-3L,	/* -0x1226e355e6c23c8f5a5a0f386cb4d.0p-119L */
68 T13 =  3.59212803657248101358314398220822722e-3L,	/*  0x1d6d3d0e157ddfb403ad3637442c6.0p-121L */
69 T15 = -1.45583438705131796512568010348874662e-3L;	/* -0x17da36452b75e150c44cc34253b34.0p-122L */
70 static const double
71 T17 =  5.9002744094556621e-4,		/*  0x1355824803668e.0p-63 */
72 T19 = -2.3912911424260516e-4,		/* -0x1f57d7734c8dde.0p-65 */
73 T21 =  9.6915379535512898e-5,		/*  0x1967e18ad6a6ca.0p-66 */
74 T23 = -3.9278322983156353e-5,		/* -0x1497d8e6b75729.0p-67 */
75 T25 =  1.5918887220143869e-5,		/*  0x10b1319998cafa.0p-68 */
76 T27 = -6.4514295231630956e-6,		/* -0x1b0f2b71b218eb.0p-70 */
77 T29 =  2.6120754043964365e-6,		/*  0x15e963a3cf3a39.0p-71 */
78 T31 = -1.0407567231003314e-6,		/* -0x1176041e656869.0p-72 */
79 T33 =  3.4744117554063574e-7;		/*  0x1750fe732cab9c.0p-74 */
80 #endif /* LDBL_MANT_DIG == 64 */
81 
82 static inline long double
83 divl(long double a, long double b, long double c, long double d,
84     long double e, long double f)
85 {
86 	long double inv, r;
87 	float fr, fw;
88 
89 	_2sumF(a, c);
90 	b = b + c;
91 	_2sumF(d, f);
92 	e = e + f;
93 
94 	inv = 1 / (d + e);
95 
96 	r = (a + b) * inv;
97 	fr = r;
98 	r = fr;
99 
100 	fw = d + e;
101 	e = d - fw + e;
102 	d = fw;
103 
104 	r = r + (a - d * r + b - e * r) * inv;
105 
106 	return r;
107 }
108 
109 long double
110 tanhl(long double x)
111 {
112 	long double hi,lo,s,x2,x4,z;
113 #if LDBL_MANT_DIG == 113
114 	double dx2;
115 #endif
116 	int16_t jx,ix;
117 
118 	GET_LDBL_EXPSIGN(jx,x);
119 	ix = jx&0x7fff;
120 
121     /* x is INF or NaN */
122 	if(ix>=0x7fff) {
123 	    if (jx>=0) return one/x+one;    /* tanh(+-inf)=+-1 */
124 	    else       return one/x-one;    /* tanh(NaN) = NaN */
125 	}
126 
127 	ENTERI();
128 
129     /* |x| < 40 */
130 	if (ix < 0x4004 || fabsl(x) < 40) {	/* |x|<40 */
131 	    if (__predict_false(ix<BIAS-(LDBL_MANT_DIG+1)/2)) {	/* |x|<TINY */
132 		/* tanh(+-0) = +0; tanh(tiny) = tiny(-+) with inexact: */
133 		return (x == 0 ? x : (0x1p200 * x - x) * 0x1p-200);
134 	    }
135 	    if (ix<0x3ffd) {		/* |x|<0.25 */
136 		x2 = x*x;
137 #if LDBL_MANT_DIG == 64
138 		x4 = x2*x2;
139 		RETURNI(((T19*x2 + T17)*x4 + (T15*x2 + T13))*(x2*x*x2*x4*x4) +
140 		    ((T11*x2 + T9)*x4 + (T7*x2 + T5))*(x2*x*x2) +
141 		    T3*(x2*x) + x);
142 #elif LDBL_MANT_DIG == 113
143 		dx2 = x2;
144 #if 0
145 		RETURNI(((((((((((((((T33*dx2 + T31)*dx2 + T29)*dx2 + T27)*dx2 +
146 		    T25)*x2 + T23)*x2 + T21)*x2 + T19)*x2 + T17)*x2 +
147 		    T15)*x2 + T13)*x2 + T11)*x2 + T9)*x2 + T7)*x2 + T5)*
148 		    (x2*x*x2) +
149 		    T3*(x2*x) + x);
150 #else
151 		long double q = ((((((((((((((T33*dx2 + T31)*dx2 + T29)*dx2 + T27)*dx2 +
152 		    T25)*x2 + T23)*x2 + T21)*x2 + T19)*x2 + T17)*x2 +
153 		    T15)*x2 + T13)*x2 + T11)*x2 + T9)*x2 + T7)*x2 + T5)*
154 		    (x2*x*x2);
155 		RETURNI(q + T3*(x2*x) + x);
156 #endif
157 #endif
158 	    }
159 	    k_hexpl(2*fabsl(x), &hi, &lo);
160 	    if (ix<0x4001 && fabsl(x) < 1.5)	/* |x|<1.5 */
161 		z = divl(hi, lo, -0.5, hi, lo, 0.5);
162 	    else
163 		z = one - one/(lo+0.5+hi);
164     /* |x| >= 40, return +-1 */
165 	} else {
166 	    z = one - tiny;		/* raise inexact flag */
167 	}
168 	s = 1;
169 	if (jx<0) s = -1;
170 	RETURNI(s*z);
171 }
172