xref: /illumos-gate/usr/src/lib/libc/i386/fp/_X_cplx_div.c (revision 581cede61ac9c14d8d4ea452562a567189eead78)
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 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * _X_cplx_div(z, w) returns z / w with infinities handled according
31  * to C99.
32  *
33  * If z and w are both finite and w is nonzero, _X_cplx_div delivers
34  * the complex quotient q according to the usual formula: let a =
35  * Re(z), b = Im(z), c = Re(w), and d = Im(w); then q = x + I * y
36  * where x = (a * c + b * d) / r and y = (b * c - a * d) / r with
37  * r = c * c + d * d.  This implementation scales to avoid premature
38  * underflow or overflow.
39  *
40  * If z is neither NaN nor zero and w is zero, or if z is infinite
41  * and w is finite and nonzero, _X_cplx_div delivers an infinite
42  * result.  If z is finite and w is infinite, _X_cplx_div delivers
43  * a zero result.
44  *
45  * If z and w are both zero or both infinite, or if either z or w is
46  * a complex NaN, _X_cplx_div delivers NaN + I * NaN.  C99 doesn't
47  * specify these cases.
48  *
49  * This implementation can raise spurious underflow, overflow, in-
50  * valid operation, inexact, and division-by-zero exceptions.  C99
51  * allows this.
52  */
53 
54 #if !defined(i386) && !defined(__i386) && !defined(__amd64)
55 #error This code is for x86 only
56 #endif
57 
58 static union {
59 	int	i;
60 	float	f;
61 } inf = {
62 	0x7f800000
63 };
64 
65 /*
66  * Return +1 if x is +Inf, -1 if x is -Inf, and 0 otherwise
67  */
68 static int
69 testinfl(long double x)
70 {
71 	union {
72 		int		i[3];
73 		long double	e;
74 	} xx;
75 
76 	xx.e = x;
77 	if ((xx.i[2] & 0x7fff) != 0x7fff || ((xx.i[1] << 1) | xx.i[0]) != 0)
78 		return (0);
79 	return (1 | ((xx.i[2] << 16) >> 31));
80 }
81 
82 long double _Complex
83 _X_cplx_div(long double _Complex z, long double _Complex w)
84 {
85 	long double _Complex	v;
86 	union {
87 		int		i[3];
88 		long double	e;
89 	} aa, bb, cc, dd, ss;
90 	long double	a, b, c, d, r;
91 	int		ea, eb, ec, ed, ez, ew, es, i, j;
92 
93 	/*
94 	 * The following is equivalent to
95 	 *
96 	 *  a = creall(*z); b = cimagl(*z);
97 	 *  c = creall(*w); d = cimagl(*w);
98 	 */
99 	a = ((long double *)&z)[0];
100 	b = ((long double *)&z)[1];
101 	c = ((long double *)&w)[0];
102 	d = ((long double *)&w)[1];
103 
104 	/* extract exponents to estimate |z| and |w| */
105 	aa.e = a;
106 	bb.e = b;
107 	ea = aa.i[2] & 0x7fff;
108 	eb = bb.i[2] & 0x7fff;
109 	ez = (ea > eb)? ea : eb;
110 
111 	cc.e = c;
112 	dd.e = d;
113 	ec = cc.i[2] & 0x7fff;
114 	ed = dd.i[2] & 0x7fff;
115 	ew = (ec > ed)? ec : ed;
116 
117 	/* check for special cases */
118 	if (ew >= 0x7fff) { /* w is inf or nan */
119 		r = 0.0f;
120 		i = testinfl(c);
121 		j = testinfl(d);
122 		if (i | j) { /* w is infinite */
123 			/*
124 			 * "factor out" infinity, being careful to preserve
125 			 * signs of finite values
126 			 */
127 			c = i? i : (((cc.i[2] << 16) < 0)? -0.0f : 0.0f);
128 			d = j? j : (((dd.i[2] << 16) < 0)? -0.0f : 0.0f);
129 			if (ez >= 0x7ffe) {
130 				/* scale to avoid overflow below */
131 				c *= 0.5f;
132 				d *= 0.5f;
133 			}
134 		}
135 		((long double *)&v)[0] = (a * c + b * d) * r;
136 		((long double *)&v)[1] = (b * c - a * d) * r;
137 		return (v);
138 	}
139 
140 	if (ew == 0 && (cc.i[1] | cc.i[0] | dd.i[1] | dd.i[0]) == 0) {
141 		/* w is zero; multiply z by 1/Re(w) - I * Im(w) */
142 		c = 1.0f / c;
143 		i = testinfl(a);
144 		j = testinfl(b);
145 		if (i | j) { /* z is infinite */
146 			a = i;
147 			b = j;
148 		}
149 		((long double *)&v)[0] = a * c + b * d;
150 		((long double *)&v)[1] = b * c - a * d;
151 		return (v);
152 	}
153 
154 	if (ez >= 0x7fff) { /* z is inf or nan */
155 		i = testinfl(a);
156 		j = testinfl(b);
157 		if (i | j) { /* z is infinite */
158 			a = i;
159 			b = j;
160 			r = inf.f;
161 		}
162 		((long double *)&v)[0] = a * c + b * d;
163 		((long double *)&v)[1] = b * c - a * d;
164 		return (v);
165 	}
166 
167 	/*
168 	 * Scale c and d to compute 1/|w|^2 and the real and imaginary
169 	 * parts of the quotient.
170 	 */
171 	es = ((ew >> 2) - ew) + 0x6ffd;
172 	if (ez < 0x0086) { /* |z| < 2^-16249 */
173 		if (((ew - 0x3efe) | (0x4083 - ew)) >= 0)
174 			es = ((0x4083 - ew) >> 1) + 0x3fff;
175 	}
176 	ss.i[2] = es;
177 	ss.i[1] = 0x80000000;
178 	ss.i[0] = 0;
179 
180 	c *= ss.e;
181 	d *= ss.e;
182 	r = 1.0f / (c * c + d * d);
183 
184 	c *= ss.e;
185 	d *= ss.e;
186 
187 	((long double *)&v)[0] = (a * c + b * d) * r;
188 	((long double *)&v)[1] = (b * c - a * d) * r;
189 	return (v);
190 }
191