xref: /illumos-gate/usr/src/lib/libm/common/C/hypot.c (revision ddc0e0b5)
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 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __hypot = hypot
31 
32 /* INDENT OFF */
33 /*
34  * Hypot(x, y)
35  * by K.C. Ng for SUN 4.0 libm, updated 3/11/2003.
36  * Method :
37  * A. When rounding is rounded-to-nearest:
38  *	If z = x * x + y * y has error less than sqrt(2) / 2 ulp than
39  *	sqrt(z) has error less than 1 ulp.
40  *	So, compute sqrt(x*x+y*y) with some care as follows:
41  *	Assume x > y > 0;
42  *	1. Check whether save and set rounding to round-to-nearest
43  *	2. if x > 2y  use
44  *		xh*xh+(y*y+((x-xh)*(x+xh))) for x*x+y*y
45  *	where xh = x with lower 32 bits cleared;  else
46  *	3. if x <= 2y use
47  *		x2h*yh+((x-y)*(x-y)+(x2h*(y-yh)+(x2-x2h)*y))
48  *	where x2 = 2*x, x2h = 2x with lower 32 bits cleared, yh = y with
49  *	lower 32 bits chopped.
50  *
51  * B. When rounding is not rounded-to-nearest:
52  *	The following (magic) formula will yield an error less than 1 ulp.
53  *	z = sqrt(x * x + y * y)
54  *		hypot(x, y) = x + (y / ((x + z) / y))
55  *
56  * NOTE: DO NOT remove parenthsis!
57  *
58  * Special cases:
59  *	hypot(x, y) is INF if x or y is +INF or -INF; else
60  *	hypot(x, y) is NAN if x or y is NAN.
61  *
62  * Accuracy:
63  * 	hypot(x, y) returns sqrt(x^2+y^2) with error less than 1 ulps
64  *	(units in the last place)
65  */
66 
67 #include "libm.h"
68 
69 static const double
70 	zero = 0.0,
71 	onep1u = 1.00000000000000022204e+00,	/* 0x3ff00000 1 = 1+2**-52 */
72 	twom53 = 1.11022302462515654042e-16,	/* 0x3ca00000 0 = 2**-53 */
73 	twom768 = 6.441148769597133308e-232,	/* 2^-768 */
74 	two768  = 1.552518092300708935e+231;	/* 2^768 */
75 
76 /* INDENT ON */
77 
78 double
hypot(double x,double y)79 hypot(double x, double y) {
80 	double xh, yh, w, ax, ay;
81 	int i, j, nx, ny, ix, iy, iscale = 0;
82 	unsigned lx, ly;
83 
84 	ix = ((int *) &x)[HIWORD] & ~0x80000000;
85 	lx = ((int *) &x)[LOWORD];
86 	iy = ((int *) &y)[HIWORD] & ~0x80000000;
87 	ly = ((int *) &y)[LOWORD];
88 /*
89  * Force ax = |x| ~>~ ay = |y|
90  */
91 	if (iy > ix) {
92 		ax = fabs(y);
93 		ay = fabs(x);
94 		i = ix;
95 		ix = iy;
96 		iy = i;
97 		i = lx;
98 		lx = ly;
99 		ly = i;
100 	} else {
101 		ax = fabs(x);
102 		ay = fabs(y);
103 	}
104 	nx = ix >> 20;
105 	ny = iy >> 20;
106 	j  = nx - ny;
107 /*
108  * x >= 2^500 (x*x or y*y may overflow)
109  */
110 	if (nx >= 0x5f3) {
111 		if (nx == 0x7ff) {	/* inf or NaN, signal of sNaN */
112 			if (((ix - 0x7ff00000) | lx) == 0)
113 				return (ax == ay ? ay : ax);
114 			else if (((iy - 0x7ff00000) | ly) == 0)
115 				return (ay == ax ? ax : ay);
116 			else
117 				return (ax * ay);	/* + -> * for Cheetah */
118 		} else if (j > 32) {	/* x >> y */
119 			if (j <= 53)
120 				ay *= twom53;
121 			ax += ay;
122 			if (((int *) &ax)[HIWORD] == 0x7ff00000)
123 				ax = _SVID_libm_err(x, y, 4);
124 			return (ax);
125 		}
126 		ax *= twom768;
127 		ay *= twom768;
128 		iscale = 2;
129 		ix -= 768 << 20;
130 		iy -= 768 << 20;
131 	}
132 /*
133  * y < 2^-450 (x*x or y*y may underflow)
134  */
135 	else if (ny < 0x23d) {
136 		if ((ix | lx) == 0)
137 			return (ay);
138 		if ((iy | ly) == 0)
139 			return (ax);
140 		if (j > 53) 		/* x >> y */
141 			return (ax + ay);
142 		iscale = 1;
143 		ax *= two768;
144 		ay *= two768;
145 		if (nx == 0) {
146 			if (ax == zero)	/* guard subnormal flush to zero */
147 				return (ax);
148 			ix = ((int *) &ax)[HIWORD];
149 		} else
150 			ix += 768 << 20;
151 		if (ny == 0) {
152 			if (ay == zero)	/* guard subnormal flush to zero */
153 				return (ax * twom768);
154 			iy = ((int *) &ay)[HIWORD];
155 		} else
156 			iy += 768 << 20;
157 		j = (ix >> 20) - (iy >> 20);
158 		if (j > 32) {		/* x >> y */
159 			if (j <= 53)
160 				ay *= twom53;
161 			return ((ax + ay) * twom768);
162 		}
163 	} else if (j > 32) {		/* x >> y */
164 		if (j <= 53)
165 			ay *= twom53;
166 		return (ax + ay);
167 	}
168 /*
169  * Medium range ax and ay with max{|ax/ay|,|ay/ax|} bounded by 2^32
170  * First check rounding mode by comparing onep1u*onep1u with onep1u+twom53.
171  * Make sure the computation is done at run-time.
172  */
173 	if (((lx | ly) << 5) == 0) {
174 		ay = ay * ay;
175 		ax += ay / (ax + sqrt(ax * ax + ay));
176 	} else
177 	if (onep1u * onep1u != onep1u + twom53) {
178 	/* round-to-zero, positive, negative mode */
179 	/* magic formula with less than an ulp error */
180 		w = sqrt(ax * ax + ay * ay);
181 		ax += ay / ((ax + w) / ay);
182 	} else {
183 	/* round-to-nearest mode */
184 		w = ax - ay;
185 		if (w > ay) {
186 			((int *) &xh)[HIWORD] = ix;
187 			((int *) &xh)[LOWORD] = 0;
188 			ay = ay * ay + (ax - xh) * (ax + xh);
189 			ax = sqrt(xh * xh + ay);
190 		} else {
191 			ax = ax + ax;
192 			((int *) &xh)[HIWORD] = ix + 0x00100000;
193 			((int *) &xh)[LOWORD] = 0;
194 			((int *) &yh)[HIWORD] = iy;
195 			((int *) &yh)[LOWORD] = 0;
196 			ay = w * w + ((ax - xh) * yh + (ay - yh) * ax);
197 			ax = sqrt(xh * yh + ay);
198 		}
199 	}
200 	if (iscale > 0) {
201 		if (iscale == 1)
202 			ax *= twom768;
203 		else {
204 			ax *= two768;	/* must generate side effect here */
205 			if (((int *) &ax)[HIWORD] == 0x7ff00000)
206 				ax = _SVID_libm_err(x, y, 4);
207 		}
208 	}
209 	return (ax);
210 }
211