xref: /illumos-gate/usr/src/lib/libm/common/complex/cabs.c (revision a9ba5504895c94d55fd9eb409fdb92f2385c9515)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
23  */
24 /*
25  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
26  * Use is subject to license terms.
27  */
28 
29 #pragma weak __cabs = cabs
30 
31 #include <math.h>
32 #include "complex_wrapper.h"
33 
34 /*
35  * If C were the only standard we cared about, cabs could just call
36  * hypot.  Unfortunately, various other standards say that hypot must
37  * call matherr and/or set errno to ERANGE when the result overflows.
38  * Since cabs should do neither of these things, we have to either
39  * make hypot a wrapper on another internal function or duplicate
40  * the hypot implementation here.  I've chosen to do the latter.
41  */
42 
43 static const double
44 	zero = 0.0,
45 	onep1u = 1.00000000000000022204e+00,	/* 0x3ff00000 1 = 1+2**-52 */
46 	twom53 = 1.11022302462515654042e-16,	/* 0x3ca00000 0 = 2**-53 */
47 	twom768 = 6.441148769597133308e-232,	/* 2^-768 */
48 	two768  = 1.552518092300708935e+231;	/* 2^768 */
49 
50 double
51 cabs(dcomplex z)
52 {
53 	double		x, y, xh, yh, w, ax, ay;
54 	int		i, j, nx, ny, ix, iy, iscale = 0;
55 	unsigned	lx, ly;
56 
57 	x = D_RE(z);
58 	y = D_IM(z);
59 
60 	ix = ((int *)&x)[HIWORD] & ~0x80000000;
61 	lx = ((int *)&x)[LOWORD];
62 	iy = ((int *)&y)[HIWORD] & ~0x80000000;
63 	ly = ((int *)&y)[LOWORD];
64 
65 	/* force ax = |x| ~>~ ay = |y| */
66 	if (iy > ix) {
67 		ax = fabs(y);
68 		ay = fabs(x);
69 		i = ix;
70 		ix = iy;
71 		iy = i;
72 		i = lx;
73 		lx = ly;
74 		ly = i;
75 	} else {
76 		ax = fabs(x);
77 		ay = fabs(y);
78 	}
79 	nx = ix >> 20;
80 	ny = iy >> 20;
81 	j  = nx - ny;
82 
83 	if (nx >= 0x5f3) {
84 		/* x >= 2^500 (x*x or y*y may overflow) */
85 		if (nx == 0x7ff) {
86 			/* inf or NaN, signal of sNaN */
87 			if (((ix - 0x7ff00000) | lx) == 0)
88 				return ((ax == ay)? ay : ax);
89 			else if (((iy - 0x7ff00000) | ly) == 0)
90 				return ((ay == ax)? ax : ay);
91 			else
92 				return (ax * ay);
93 		} else if (j > 32) {
94 			/* x >> y */
95 			if (j <= 53)
96 				ay *= twom53;
97 			ax += ay;
98 			return (ax);
99 		}
100 		ax *= twom768;
101 		ay *= twom768;
102 		iscale = 2;
103 		ix -= 768 << 20;
104 		iy -= 768 << 20;
105 	} else if (ny < 0x23d) {
106 		/* y < 2^-450 (x*x or y*y may underflow) */
107 		if ((ix | lx) == 0)
108 			return (ay);
109 		if ((iy | ly) == 0)
110 			return (ax);
111 		if (j > 53) 		/* x >> y */
112 			return (ax + ay);
113 		iscale = 1;
114 		ax *= two768;
115 		ay *= two768;
116 		if (nx == 0) {
117 			if (ax == zero)	/* guard subnormal flush to zero */
118 				return (ax);
119 			ix = ((int *)&ax)[HIWORD];
120 		} else {
121 			ix += 768 << 20;
122 		}
123 		if (ny == 0) {
124 			if (ay == zero)	/* guard subnormal flush to zero */
125 				return (ax * twom768);
126 			iy = ((int *)&ay)[HIWORD];
127 		} else {
128 			iy += 768 << 20;
129 		}
130 		j = (ix >> 20) - (iy >> 20);
131 		if (j > 32) {
132 			/* x >> y */
133 			if (j <= 53)
134 				ay *= twom53;
135 			return ((ax + ay) * twom768);
136 		}
137 	} else if (j > 32) {
138 		/* x >> y */
139 		if (j <= 53)
140 			ay *= twom53;
141 		return (ax + ay);
142 	}
143 
144 	/*
145 	 * Medium range ax and ay with max{|ax/ay|,|ay/ax|} bounded by 2^32.
146 	 * First check rounding mode by comparing onep1u*onep1u with onep1u
147 	 * + twom53.  Make sure the computation is done at run-time.
148 	 */
149 	if (((lx | ly) << 5) == 0) {
150 		ay = ay * ay;
151 		ax += ay / (ax + sqrt(ax * ax + ay));
152 	} else if (onep1u * onep1u != onep1u + twom53) {
153 		/* round-to-zero, positive, negative mode */
154 		/* magic formula with less than an ulp error */
155 		w = sqrt(ax * ax + ay * ay);
156 		ax += ay / ((ax + w) / ay);
157 	} else {
158 		/* round-to-nearest mode */
159 		w = ax - ay;
160 		if (w > ay) {
161 			((int *)&xh)[HIWORD] = ix;
162 			((int *)&xh)[LOWORD] = 0;
163 			ay = ay * ay + (ax - xh) * (ax + xh);
164 			ax = sqrt(xh * xh + ay);
165 		} else {
166 			ax = ax + ax;
167 			((int *)&xh)[HIWORD] = ix + 0x00100000;
168 			((int *)&xh)[LOWORD] = 0;
169 			((int *)&yh)[HIWORD] = iy;
170 			((int *)&yh)[LOWORD] = 0;
171 			ay = w * w + ((ax - xh) * yh + (ay - yh) * ax);
172 			ax = sqrt(xh * yh + ay);
173 		}
174 	}
175 	if (iscale > 0) {
176 		if (iscale == 1)
177 			ax *= twom768;
178 		else
179 			ax *= two768;	/* must generate side effect here */
180 	}
181 	return (ax);
182 }
183