xref: /netbsd-src/lib/libm/src/s_tanpi.c (revision cfe182f36bde4c4d81e1607954ce22a67cf35d7a)
1 /*-
2  * Copyright (c) 2017, 2023 Steven G. Kargl
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 /**
28  * tanpi(x) computes tan(pi*x) without multiplication by pi (almost).  First,
29  * note that tanpi(-x) = -tanpi(x), so the algorithm considers only |x| and
30  * includes reflection symmetry by considering the sign of x on output.  The
31  * method used depends on the magnitude of x.
32  *
33  * 1. For small |x|, tanpi(x) = pi * x where a sloppy threshold is used.  The
34  *    threshold is |x| < 0x1pN with N = -(P/2+M).  P is the precision of the
35  *    floating-point type and M = 2 to 4.  To achieve high accuracy, pi is
36  *    decomposed into high and low parts with the high part containing a
37  *    number of trailing zero bits.  x is also split into high and low parts.
38  *
39  * 2. For |x| < 1, argument reduction is not required and tanpi(x) is
40  *    computed by a direct call to a kernel, which uses the kernel for
41  *    tan(x).  See below.
42  *
43  * 3. For 1 <= |x| < 0x1p(P-1), argument reduction is required where
44  *    |x| = j0 + r with j0 an integer and the remainder r satisfies
45  *    0 <= r < 1.  With the given domain, a simplified inline floor(x)
46  *    is used.  Also, note the following identity
47  *
48  *                                   tan(pi*j0) + tan(pi*r)
49  *    tanpi(x) = tan(pi*(j0+r)) = ---------------------------- = tanpi(r)
50  *                                 1 - tan(pi*j0) * tan(pi*r)
51  *
52  *    So, after argument reduction, the kernel is again invoked.
53  *
54  * 4. For |x| >= 0x1p(P-1), |x| is integral and tanpi(x) = copysign(0,x).
55  *
56  * 5. Special cases:
57  *
58  *    tanpi(+-0) = +-0
59  *    tanpi(n) = +0 for positive even and negative odd integer n.
60  *    tanpi(n) = -0 for positive odd and negative even integer n.
61  *    tanpi(+-n+1/4) = +-1, for positive integers n.
62  *    tanpi(n+1/2) = +inf and raises the FE_DIVBYZERO exception for
63  *    even integers n.
64  *    tanpi(n+1/2) = -inf and raises the FE_DIVBYZERO exception for
65  *    odd integers n.
66  *    tanpi(+-inf) = NaN and raises the FE_INVALID exception.
67  *    tanpi(nan) = NaN and raises the FE_INVALID exception.
68  */
69 
70 #include <sys/cdefs.h>
71 
72 #include "namespace.h"
73 __weak_alias(tanpi, _tanpi)
74 #include <float.h>
75 #include "math.h"
76 #include "math_private.h"
77 
78 static const double
79 pi_hi =  3.1415926814079285e+00,	/* 0x400921fb 0x58000000 */
80 pi_lo = -2.7818135228334233e-08;	/* 0xbe5dde97 0x3dcb3b3a */
81 
82 /*
83  * The kernel for tanpi(x) multiplies x by an 80-bit approximation of
84  * pi, where the hi and lo parts are used with with kernel for tan(x).
85  */
86 static inline double
__kernel_tanpi(double x)87 __kernel_tanpi(double x)
88 {
89 	double_t hi, lo, t;
90 
91 	if (x < 0.25) {
92 		hi = (float)x;
93 		lo = x - hi;
94 		lo = lo * (pi_lo + pi_hi) + hi * pi_lo;
95 		hi *= pi_hi;
96 		_2sumF(hi, lo);
97 		t = __kernel_tan(hi, lo, 1);
98 	} else if (x > 0.25) {
99 		x = 0.5 - x;
100 		hi = (float)x;
101 		lo = x - hi;
102 		lo = lo * (pi_lo + pi_hi) + hi * pi_lo;
103 		hi *= pi_hi;
104 		_2sumF(hi, lo);
105 		t = - __kernel_tan(hi, lo, -1);
106 	} else
107 		t = 1;
108 
109 	return (t);
110 }
111 
112 static volatile const double vzero = 0;
113 
114 double
tanpi(double x)115 tanpi(double x)
116 {
117 	double ax, hi, lo, odd, t;
118 	uint32_t hx, ix, j0, lx;
119 
120 	EXTRACT_WORDS(hx, lx, x);
121 	ix = hx & 0x7fffffff;
122 	INSERT_WORDS(ax, ix, lx);
123 
124 	if (ix < 0x3ff00000) {			/* |x| < 1 */
125 		if (ix < 0x3fe00000) {		/* |x| < 0.5 */
126 			if (ix < 0x3e200000) {	/* |x| < 0x1p-29 */
127 				if (x == 0)
128 					return (x);
129 				/*
130 				 * To avoid issues with subnormal values,
131 				 * scale the computation and rescale on
132 				 * return.
133 				 */
134 				INSERT_WORDS(hi, hx, 0);
135 				hi *= 0x1p53;
136 				lo = x * 0x1p53 - hi;
137 				t = (pi_lo + pi_hi) * lo + pi_lo * hi +
138 				    pi_hi * hi;
139 				return (t * 0x1p-53);
140 			}
141 			t = __kernel_tanpi(ax);
142 		} else if (ax == 0.5)
143 			t = 1 / vzero;
144 		else
145 			t = - __kernel_tanpi(1 - ax);
146 		return ((hx & 0x80000000) ? -t : t);
147 	}
148 
149 	if (ix < 0x43300000) {		/* 1 <= |x| < 0x1p52 */
150 		FFLOOR(x, j0, ix, lx);	/* Integer part of ax. */
151 		odd = (uint64_t)x & 1 ? -1 : 1;
152 		ax -= x;
153 		EXTRACT_WORDS(ix, lx, ax);
154 
155 		if (ix < 0x3fe00000)		/* |x| < 0.5 */
156 			t = ix == 0 ? copysign(0, odd) : __kernel_tanpi(ax);
157 		else if (ax == 0.5)
158 			t = odd / vzero;
159 		else
160 			t = - __kernel_tanpi(1 - ax);
161 
162 		return ((hx & 0x80000000) ? -t : t);
163 	}
164 
165 	/* x = +-inf or nan. */
166 	if (ix >= 0x7ff00000)
167 		return (vzero / vzero);
168 
169 	/*
170 	 * For 0x1p52 <= |x| < 0x1p53 need to determine if x is an even
171 	 * or odd integer to set t = +0 or -0.
172 	 * For |x| >= 0x1p54, it is always an even integer, so t = 0.
173 	 */
174 	t = ix >= 0x43400000 ? 0 : (copysign(0, (lx & 1) ? -1 : 1));
175 	return ((hx & 0x80000000) ? -t : t);
176 }
177