xref: /freebsd/lib/msun/ld80/k_cosl.c (revision ca987d4641cdcd7f27e153db17c5bf064934faf5)
1 /* From: @(#)k_cos.c 1.3 95/01/18 */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
6  *
7  * Developed at SunSoft, a Sun Microsystems, Inc. business.
8  * Permission to use, copy, modify, and distribute this
9  * software is freely granted, provided that this notice
10  * is preserved.
11  * ====================================================
12  */
13 
14 #include <sys/cdefs.h>
15 __FBSDID("$FreeBSD$");
16 
17 /*
18  * ld80 version of k_cos.c.  See ../src/k_cos.c for most comments.
19  */
20 
21 #include "math_private.h"
22 
23 /*
24  * Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]:
25  * |cos(x) - c(x)| < 2**-75.1
26  *
27  * The coefficients of c(x) were generated by a pari-gp script using
28  * a Remez algorithm that searches for the best higher coefficients
29  * after rounding leading coefficients to a specified precision.
30  *
31  * Simpler methods like Chebyshev or basic Remez barely suffice for
32  * cos() in 64-bit precision, because we want the coefficient of x^2
33  * to be precisely -0.5 so that multiplying by it is exact, and plain
34  * rounding of the coefficients of a good polynomial approximation only
35  * gives this up to about 64-bit precision.  Plain rounding also gives
36  * a mediocre approximation for the coefficient of x^4, but a rounding
37  * error of 0.5 ulps for this coefficient would only contribute ~0.01
38  * ulps to the final error, so this is unimportant.  Rounding errors in
39  * higher coefficients are even less important.
40  *
41  * In fact, coefficients above the x^4 one only need to have 53-bit
42  * precision, and this is more efficient.  We get this optimization
43  * almost for free from the complications needed to search for the best
44  * higher coefficients.
45  */
46 static const double
47 one = 1.0;
48 
49 #if defined(__amd64__) || defined(__i386__)
50 /* Long double constants are slow on these arches, and broken on i386. */
51 static const volatile double
52 C1hi = 0.041666666666666664,		/*  0x15555555555555.0p-57 */
53 C1lo = 2.2598839032744733e-18;		/*  0x14d80000000000.0p-111 */
54 #define	C1	((long double)C1hi + C1lo)
55 #else
56 static const long double
57 C1 =  0.0416666666666666666136L;	/*  0xaaaaaaaaaaaaaa9b.0p-68 */
58 #endif
59 
60 static const double
61 C2 = -0.0013888888888888874,		/* -0x16c16c16c16c10.0p-62 */
62 C3 =  0.000024801587301571716,		/*  0x1a01a01a018e22.0p-68 */
63 C4 = -0.00000027557319215507120,	/* -0x127e4fb7602f22.0p-74 */
64 C5 =  0.0000000020876754400407278,	/*  0x11eed8caaeccf1.0p-81 */
65 C6 = -1.1470297442401303e-11,		/* -0x19393412bd1529.0p-89 */
66 C7 =  4.7383039476436467e-14;		/*  0x1aac9d9af5c43e.0p-97 */
67 
68 long double
69 __kernel_cosl(long double x, long double y)
70 {
71 	long double hz,z,r,w;
72 
73 	z  = x*x;
74 	r  = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7))))));
75 	hz = 0.5*z;
76 	w  = one-hz;
77 	return w + (((one-w)-hz) + (z*r-x*y));
78 }
79