xref: /illumos-gate/usr/src/lib/libm/common/LD/hypotl.c (revision 9e573dcc)
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 #if defined(ELFOBJ)
31 #pragma weak hypotl = __hypotl
32 #endif
33 
34 /*
35  * hypotl(x,y)
36  * Method :
37  *	If z=x*x+y*y has error less than sqrt(2)/2 ulp than sqrt(z) has
38  *	error less than 1 ulp.
39  *	So, compute sqrt(x*x+y*y) with some care as follows:
40  *	Assume x>y>0;
41  *	1. save and set rounding to round-to-nearest
42  *	2. if x > 2y  use
43  *		x1*x1+(y*y+(x2*(x+x2))) for x*x+y*y
44  *	where x1 = x with lower 32 bits cleared, x2 = x-x1; else
45  *	3. if x <= 2y use
46  *		t1*y1+((x-y)*(x-y)+(t1*y2+t2*y))
47  *	where t1 = 2x with lower 64 bits cleared, t2 = 2x-t1, y1= y with
48  *	lower 32 bits cleared, y2 = y-y1.
49  *
50  *	NOTE: DO NOT remove parenthsis!
51  *
52  * Special cases:
53  *	hypot(x,y) is INF if x or y is +INF or -INF; else
54  *	hypot(x,y) is NAN if x or y is NAN.
55  *
56  * Accuracy:
57  * 	hypot(x,y) returns sqrt(x^2+y^2) with error less than 1 ulps (units
58  *	in the last place)
59  */
60 
61 #include "libm.h"
62 
63 #if defined(__x86)
64 extern enum fp_direction_type __swap87RD(enum fp_direction_type);
65 
66 #define	k	0x7fff
67 
68 long double
69 hypotl(long double x, long double y) {
70 	long double t1, t2, y1, y2, w;
71 	int *px = (int *) &x, *py = (int *) &y;
72 	int *pt1 = (int *) &t1, *py1 = (int *) &y1;
73 	enum fp_direction_type rd;
74 	int j, nx, ny, nz;
75 
76 	px[2] &= 0x7fff;	/* clear sign bit and padding bits of x and y */
77 	py[2] &= 0x7fff;
78 	nx = px[2];		/* biased exponent of x and y */
79 	ny = py[2];
80 	if (ny > nx) {
81 		w = x;
82 		x = y;
83 		y = w;
84 		nz = ny;
85 		ny = nx;
86 		nx = nz;
87 	}			/* force nx >= ny */
88 	if (nx - ny >= 66)
89 		return (x + y);	/* x / y >= 2**65 */
90 	if (nx < 0x5ff3 && ny > 0x205b) {	/* medium x,y */
91 		/* save and set RD to Rounding to nearest */
92 		rd = __swap87RD(fp_nearest);
93 		w = x - y;
94 		if (w > y) {
95 			pt1[2] = px[2];
96 			pt1[1] = px[1];
97 			pt1[0] = 0;
98 			t2 = x - t1;
99 			x = sqrtl(t1 * t1 - (y * (-y) - t2 * (x + t1)));
100 		} else {
101 			x += x;
102 			py1[2] = py[2];
103 			py1[1] = py[1];
104 			py1[0] = 0;
105 			y2 = y - y1;
106 			pt1[2] = px[2];
107 			pt1[1] = px[1];
108 			pt1[0] = 0;
109 			t2 = x - t1;
110 			x = sqrtl(t1 * y1 - (w * (-w) - (t2 * y1 + y2 * x)));
111 		}
112 		if (rd != fp_nearest)
113 			__swap87RD(rd);	/* restore rounding mode */
114 		return (x);
115 	} else {
116 		if (nx == k || ny == k) {	/* x or y is INF or NaN */
117 			/* since nx >= ny; nx is always k within this block */
118 			if (px[1] == 0x80000000 && px[0] == 0)
119 				return (x);
120 			else if (ny == k && py[1] == 0x80000000 && py[0] == 0)
121 				return (y);
122 			else
123 				return (x + y);
124 		}
125 		if (ny == 0) {
126 			if (y == 0.L || x == 0.L)
127 				return (x + y);
128 			pt1[2] = 0x3fff + 16381;
129 			pt1[1] = 0x80000000;
130 			pt1[0] = 0;
131 			py1[2] = 0x3fff - 16381;
132 			py1[1] = 0x80000000;
133 			py1[0] = 0;
134 			x *= t1;
135 			y *= t1;
136 			return (y1 * hypotl(x, y));
137 		}
138 		j = nx - 0x3fff;
139 		px[2] -= j;
140 		py[2] -= j;
141 		pt1[2] = nx;
142 		pt1[1] = 0x80000000;
143 		pt1[0] = 0;
144 		return (t1 * hypotl(x, y));
145 	}
146 }
147 #endif
148