xref: /freebsd/lib/msun/src/s_clogl.c (revision 02e9120893770924227138ba49df1edb3896112a)
1 /*-
2  * Copyright (c) 2013 Bruce D. Evans
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice unmodified, this list of conditions, and the following
10  *    disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
16  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
17  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
18  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
19  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
20  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
21  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
22  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
23  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
24  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
25  */
26 
27 #include <sys/cdefs.h>
28 #include <complex.h>
29 #include <float.h>
30 #ifdef __i386__
31 #include <ieeefp.h>
32 #endif
33 
34 #include "fpmath.h"
35 #include "math.h"
36 #include "math_private.h"
37 
38 #define	MANT_DIG	LDBL_MANT_DIG
39 #define	MAX_EXP		LDBL_MAX_EXP
40 #define	MIN_EXP		LDBL_MIN_EXP
41 
42 static const double
43 ln2_hi = 6.9314718055829871e-1;		/*  0x162e42fefa0000.0p-53 */
44 
45 #if LDBL_MANT_DIG == 64
46 #define	MULT_REDUX	0x1p32		/* exponent MANT_DIG / 2 rounded up */
47 static const double
48 ln2l_lo = 1.6465949582897082e-12;	/*  0x1cf79abc9e3b3a.0p-92 */
49 #elif LDBL_MANT_DIG == 113
50 #define	MULT_REDUX	0x1p57
51 static const long double
52 ln2l_lo = 1.64659495828970812809844307550013433e-12L;	/*  0x1cf79abc9e3b39803f2f6af40f343.0p-152L */
53 #else
54 #error "Unsupported long double format"
55 #endif
56 
57 long double complex
58 clogl(long double complex z)
59 {
60 	long double ax, ax2h, ax2l, axh, axl, ay, ay2h, ay2l, ayh, ayl;
61 	long double sh, sl, t;
62 	long double x, y, v;
63 	uint16_t hax, hay;
64 	int kx, ky;
65 
66 	ENTERIT(long double complex);
67 
68 	x = creall(z);
69 	y = cimagl(z);
70 	v = atan2l(y, x);
71 
72 	ax = fabsl(x);
73 	ay = fabsl(y);
74 	if (ax < ay) {
75 		t = ax;
76 		ax = ay;
77 		ay = t;
78 	}
79 
80 	GET_LDBL_EXPSIGN(hax, ax);
81 	kx = hax - 16383;
82 	GET_LDBL_EXPSIGN(hay, ay);
83 	ky = hay - 16383;
84 
85 	/* Handle NaNs and Infs using the general formula. */
86 	if (kx == MAX_EXP || ky == MAX_EXP)
87 		RETURNI(CMPLXL(logl(hypotl(x, y)), v));
88 
89 	/* Avoid spurious underflow, and reduce inaccuracies when ax is 1. */
90 	if (ax == 1) {
91 		if (ky < (MIN_EXP - 1) / 2)
92 			RETURNI(CMPLXL((ay / 2) * ay, v));
93 		RETURNI(CMPLXL(log1pl(ay * ay) / 2, v));
94 	}
95 
96 	/* Avoid underflow when ax is not small.  Also handle zero args. */
97 	if (kx - ky > MANT_DIG || ay == 0)
98 		RETURNI(CMPLXL(logl(ax), v));
99 
100 	/* Avoid overflow. */
101 	if (kx >= MAX_EXP - 1)
102 		RETURNI(CMPLXL(logl(hypotl(x * 0x1p-16382L, y * 0x1p-16382L)) +
103 		    (MAX_EXP - 2) * ln2l_lo + (MAX_EXP - 2) * ln2_hi, v));
104 	if (kx >= (MAX_EXP - 1) / 2)
105 		RETURNI(CMPLXL(logl(hypotl(x, y)), v));
106 
107 	/* Reduce inaccuracies and avoid underflow when ax is denormal. */
108 	if (kx <= MIN_EXP - 2)
109 		RETURNI(CMPLXL(logl(hypotl(x * 0x1p16383L, y * 0x1p16383L)) +
110 		    (MIN_EXP - 2) * ln2l_lo + (MIN_EXP - 2) * ln2_hi, v));
111 
112 	/* Avoid remaining underflows (when ax is small but not denormal). */
113 	if (ky < (MIN_EXP - 1) / 2 + MANT_DIG)
114 		RETURNI(CMPLXL(logl(hypotl(x, y)), v));
115 
116 	/* Calculate ax*ax and ay*ay exactly using Dekker's algorithm. */
117 	t = (long double)(ax * (MULT_REDUX + 1));
118 	axh = (long double)(ax - t) + t;
119 	axl = ax - axh;
120 	ax2h = ax * ax;
121 	ax2l = axh * axh - ax2h + 2 * axh * axl + axl * axl;
122 	t = (long double)(ay * (MULT_REDUX + 1));
123 	ayh = (long double)(ay - t) + t;
124 	ayl = ay - ayh;
125 	ay2h = ay * ay;
126 	ay2l = ayh * ayh - ay2h + 2 * ayh * ayl + ayl * ayl;
127 
128 	/*
129 	 * When log(|z|) is far from 1, accuracy in calculating the sum
130 	 * of the squares is not very important since log() reduces
131 	 * inaccuracies.  We depended on this to use the general
132 	 * formula when log(|z|) is very far from 1.  When log(|z|) is
133 	 * moderately far from 1, we go through the extra-precision
134 	 * calculations to reduce branches and gain a little accuracy.
135 	 *
136 	 * When |z| is near 1, we subtract 1 and use log1p() and don't
137 	 * leave it to log() to subtract 1, since we gain at least 1 bit
138 	 * of accuracy in this way.
139 	 *
140 	 * When |z| is very near 1, subtracting 1 can cancel almost
141 	 * 3*MANT_DIG bits.  We arrange that subtracting 1 is exact in
142 	 * doubled precision, and then do the rest of the calculation
143 	 * in sloppy doubled precision.  Although large cancellations
144 	 * often lose lots of accuracy, here the final result is exact
145 	 * in doubled precision if the large calculation occurs (because
146 	 * then it is exact in tripled precision and the cancellation
147 	 * removes enough bits to fit in doubled precision).  Thus the
148 	 * result is accurate in sloppy doubled precision, and the only
149 	 * significant loss of accuracy is when it is summed and passed
150 	 * to log1p().
151 	 */
152 	sh = ax2h;
153 	sl = ay2h;
154 	_2sumF(sh, sl);
155 	if (sh < 0.5 || sh >= 3)
156 		RETURNI(CMPLXL(logl(ay2l + ax2l + sl + sh) / 2, v));
157 	sh -= 1;
158 	_2sum(sh, sl);
159 	_2sum(ax2l, ay2l);
160 	/* Briggs-Kahan algorithm (except we discard the final low term): */
161 	_2sum(sh, ax2l);
162 	_2sum(sl, ay2l);
163 	t = ax2l + sl;
164 	_2sumF(sh, t);
165 	RETURNI(CMPLXL(log1pl(ay2l + t + sh) / 2, v));
166 }
167