xref: /illumos-gate/usr/src/lib/libc/i386/fp/_X_cplx_div.c (revision 1da57d55)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * _X_cplx_div(z, w) returns z / w with infinities handled according
29  * to C99.
30  *
31  * If z and w are both finite and w is nonzero, _X_cplx_div delivers
32  * the complex quotient q according to the usual formula: let a =
33  * Re(z), b = Im(z), c = Re(w), and d = Im(w); then q = x + I * y
34  * where x = (a * c + b * d) / r and y = (b * c - a * d) / r with
35  * r = c * c + d * d.  This implementation scales to avoid premature
36  * underflow or overflow.
37  *
38  * If z is neither NaN nor zero and w is zero, or if z is infinite
39  * and w is finite and nonzero, _X_cplx_div delivers an infinite
40  * result.  If z is finite and w is infinite, _X_cplx_div delivers
41  * a zero result.
42  *
43  * If z and w are both zero or both infinite, or if either z or w is
44  * a complex NaN, _X_cplx_div delivers NaN + I * NaN.  C99 doesn't
45  * specify these cases.
46  *
47  * This implementation can raise spurious underflow, overflow, in-
48  * valid operation, inexact, and division-by-zero exceptions.  C99
49  * allows this.
50  */
51 
52 #if !defined(i386) && !defined(__i386) && !defined(__amd64)
53 #error This code is for x86 only
54 #endif
55 
56 static union {
57 	int	i;
58 	float	f;
59 } inf = {
60 	0x7f800000
61 };
62 
63 /*
64  * Return +1 if x is +Inf, -1 if x is -Inf, and 0 otherwise
65  */
66 static int
testinfl(long double x)67 testinfl(long double x)
68 {
69 	union {
70 		int		i[3];
71 		long double	e;
72 	} xx;
73 
74 	xx.e = x;
75 	if ((xx.i[2] & 0x7fff) != 0x7fff || ((xx.i[1] << 1) | xx.i[0]) != 0)
76 		return (0);
77 	return (1 | ((xx.i[2] << 16) >> 31));
78 }
79 
80 long double _Complex
_X_cplx_div(long double _Complex z,long double _Complex w)81 _X_cplx_div(long double _Complex z, long double _Complex w)
82 {
83 	long double _Complex	v;
84 	union {
85 		int		i[3];
86 		long double	e;
87 	} aa, bb, cc, dd, ss;
88 	long double	a, b, c, d, r;
89 	int		ea, eb, ec, ed, ez, ew, es, i, j;
90 
91 	/*
92 	 * The following is equivalent to
93 	 *
94 	 *  a = creall(*z); b = cimagl(*z);
95 	 *  c = creall(*w); d = cimagl(*w);
96 	 */
97 	a = ((long double *)&z)[0];
98 	b = ((long double *)&z)[1];
99 	c = ((long double *)&w)[0];
100 	d = ((long double *)&w)[1];
101 
102 	/* extract exponents to estimate |z| and |w| */
103 	aa.e = a;
104 	bb.e = b;
105 	ea = aa.i[2] & 0x7fff;
106 	eb = bb.i[2] & 0x7fff;
107 	ez = (ea > eb)? ea : eb;
108 
109 	cc.e = c;
110 	dd.e = d;
111 	ec = cc.i[2] & 0x7fff;
112 	ed = dd.i[2] & 0x7fff;
113 	ew = (ec > ed)? ec : ed;
114 
115 	/* check for special cases */
116 	if (ew >= 0x7fff) { /* w is inf or nan */
117 		r = 0.0f;
118 		i = testinfl(c);
119 		j = testinfl(d);
120 		if (i | j) { /* w is infinite */
121 			/*
122 			 * "factor out" infinity, being careful to preserve
123 			 * signs of finite values
124 			 */
125 			c = i? i : (((cc.i[2] << 16) < 0)? -0.0f : 0.0f);
126 			d = j? j : (((dd.i[2] << 16) < 0)? -0.0f : 0.0f);
127 			if (ez >= 0x7ffe) {
128 				/* scale to avoid overflow below */
129 				c *= 0.5f;
130 				d *= 0.5f;
131 			}
132 		}
133 		((long double *)&v)[0] = (a * c + b * d) * r;
134 		((long double *)&v)[1] = (b * c - a * d) * r;
135 		return (v);
136 	}
137 
138 	if (ew == 0 && (cc.i[1] | cc.i[0] | dd.i[1] | dd.i[0]) == 0) {
139 		/* w is zero; multiply z by 1/Re(w) - I * Im(w) */
140 		c = 1.0f / c;
141 		i = testinfl(a);
142 		j = testinfl(b);
143 		if (i | j) { /* z is infinite */
144 			a = i;
145 			b = j;
146 		}
147 		((long double *)&v)[0] = a * c + b * d;
148 		((long double *)&v)[1] = b * c - a * d;
149 		return (v);
150 	}
151 
152 	if (ez >= 0x7fff) { /* z is inf or nan */
153 		i = testinfl(a);
154 		j = testinfl(b);
155 		if (i | j) { /* z is infinite */
156 			a = i;
157 			b = j;
158 			r = inf.f;
159 		}
160 		((long double *)&v)[0] = a * c + b * d;
161 		((long double *)&v)[1] = b * c - a * d;
162 		return (v);
163 	}
164 
165 	/*
166 	 * Scale c and d to compute 1/|w|^2 and the real and imaginary
167 	 * parts of the quotient.
168 	 */
169 	es = ((ew >> 2) - ew) + 0x6ffd;
170 	if (ez < 0x0086) { /* |z| < 2^-16249 */
171 		if (((ew - 0x3efe) | (0x4083 - ew)) >= 0)
172 			es = ((0x4083 - ew) >> 1) + 0x3fff;
173 	}
174 	ss.i[2] = es;
175 	ss.i[1] = 0x80000000;
176 	ss.i[0] = 0;
177 
178 	c *= ss.e;
179 	d *= ss.e;
180 	r = 1.0f / (c * c + d * d);
181 
182 	c *= ss.e;
183 	d *= ss.e;
184 
185 	((long double *)&v)[0] = (a * c + b * d) * r;
186 	((long double *)&v)[1] = (b * c - a * d) * r;
187 	return (v);
188 }
189