xref: /freebsd/lib/msun/src/s_csqrtl.c (revision 783d3ff6)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2007-2008 David Schultz <das@FreeBSD.ORG>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <complex.h>
30 #include <float.h>
31 #include <math.h>
32 
33 #include "math_private.h"
34 
35 /*
36  * Several thresholds require a 15-bit exponent and also the usual bias.
37  * s_logl.c and e_hypotl have less hard-coding but end up requiring the
38  * same for the exponent and more for the mantissa.
39  */
40 #if LDBL_MAX_EXP != 0x4000
41 #error "Unsupported long double format"
42 #endif
43 
44 /*
45  * Overflow must be avoided for components >= LDBL_MAX / (1 + sqrt(2)).
46  * The precise threshold is nontrivial to determine and spell, so use a
47  * lower threshold of approximaely LDBL_MAX / 4, and don't use LDBL_MAX
48  * to spell this since LDBL_MAX is broken on i386 (it overflows in 53-bit
49  * precision).
50  */
51 #define	THRESH	0x1p16382L
52 
53 long double complex
54 csqrtl(long double complex z)
55 {
56 	long double complex result;
57 	long double a, b, rx, ry, scale, t;
58 
59 	a = creall(z);
60 	b = cimagl(z);
61 
62 	/* Handle special cases. */
63 	if (z == 0)
64 		return (CMPLXL(0, b));
65 	if (isinf(b))
66 		return (CMPLXL(INFINITY, b));
67 	if (isnan(a)) {
68 		t = (b - b) / (b - b);	/* raise invalid if b is not a NaN */
69 		return (CMPLXL(a + 0.0L + t, a + 0.0L + t)); /* NaN + NaN i */
70 	}
71 	if (isinf(a)) {
72 		/*
73 		 * csqrt(inf + NaN i)  = inf +  NaN i
74 		 * csqrt(inf + y i)    = inf +  0 i
75 		 * csqrt(-inf + NaN i) = NaN +- inf i
76 		 * csqrt(-inf + y i)   = 0   +  inf i
77 		 */
78 		if (signbit(a))
79 			return (CMPLXL(fabsl(b - b), copysignl(a, b)));
80 		else
81 			return (CMPLXL(a, copysignl(b - b, b)));
82 	}
83 	if (isnan(b)) {
84 		t = (a - a) / (a - a);	/* raise invalid */
85 		return (CMPLXL(b + 0.0L + t, b + 0.0L + t)); /* NaN + NaN i */
86 	}
87 
88 	/* Scale to avoid overflow. */
89 	if (fabsl(a) >= THRESH || fabsl(b) >= THRESH) {
90 		/*
91 		 * Don't scale a or b if this might give (spurious)
92 		 * underflow.  Then the unscaled value is an equivalent
93 		 * infinitesmal (or 0).
94 		 */
95 		if (fabsl(a) >= 0x1p-16380L)
96 			a *= 0.25;
97 		if (fabsl(b) >= 0x1p-16380L)
98 			b *= 0.25;
99 		scale = 2;
100 	} else {
101 		scale = 1;
102 	}
103 
104 	/* Scale to reduce inaccuracies when both components are denormal. */
105 	if (fabsl(a) < 0x1p-16382L && fabsl(b) < 0x1p-16382L) {
106 		a *= 0x1p64;
107 		b *= 0x1p64;
108 		scale = 0x1p-32;
109 	}
110 
111 	/* Algorithm 312, CACM vol 10, Oct 1967. */
112 	if (a >= 0) {
113 		t = sqrtl((a + hypotl(a, b)) * 0.5);
114 		rx = scale * t;
115 		ry = scale * b / (2 * t);
116 	} else {
117 		t = sqrtl((-a + hypotl(a, b)) * 0.5);
118 		rx = scale * fabsl(b) / (2 * t);
119 		ry = copysignl(scale * t, b);
120 	}
121 
122 	return (CMPLXL(rx, ry));
123 }
124