xref: /freebsd/lib/msun/ld80/k_cosl.c (revision 02e9120893770924227138ba49df1edb3896112a)
1 /*
2  * ====================================================
3  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
5  *
6  * Developed at SunSoft, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 #include <sys/cdefs.h>
14 /*
15  * ld80 version of k_cos.c.  See ../src/k_cos.c for most comments.
16  */
17 
18 #include "math_private.h"
19 
20 /*
21  * Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]:
22  * |cos(x) - c(x)| < 2**-75.1
23  *
24  * The coefficients of c(x) were generated by a pari-gp script using
25  * a Remez algorithm that searches for the best higher coefficients
26  * after rounding leading coefficients to a specified precision.
27  *
28  * Simpler methods like Chebyshev or basic Remez barely suffice for
29  * cos() in 64-bit precision, because we want the coefficient of x^2
30  * to be precisely -0.5 so that multiplying by it is exact, and plain
31  * rounding of the coefficients of a good polynomial approximation only
32  * gives this up to about 64-bit precision.  Plain rounding also gives
33  * a mediocre approximation for the coefficient of x^4, but a rounding
34  * error of 0.5 ulps for this coefficient would only contribute ~0.01
35  * ulps to the final error, so this is unimportant.  Rounding errors in
36  * higher coefficients are even less important.
37  *
38  * In fact, coefficients above the x^4 one only need to have 53-bit
39  * precision, and this is more efficient.  We get this optimization
40  * almost for free from the complications needed to search for the best
41  * higher coefficients.
42  */
43 static const double
44 one = 1.0;
45 
46 #if defined(__amd64__) || defined(__i386__)
47 /* Long double constants are slow on these arches, and broken on i386. */
48 static const volatile double
49 C1hi = 0.041666666666666664,		/*  0x15555555555555.0p-57 */
50 C1lo = 2.2598839032744733e-18;		/*  0x14d80000000000.0p-111 */
51 #define	C1	((long double)C1hi + C1lo)
52 #else
53 static const long double
54 C1 =  0.0416666666666666666136L;	/*  0xaaaaaaaaaaaaaa9b.0p-68 */
55 #endif
56 
57 static const double
58 C2 = -0.0013888888888888874,		/* -0x16c16c16c16c10.0p-62 */
59 C3 =  0.000024801587301571716,		/*  0x1a01a01a018e22.0p-68 */
60 C4 = -0.00000027557319215507120,	/* -0x127e4fb7602f22.0p-74 */
61 C5 =  0.0000000020876754400407278,	/*  0x11eed8caaeccf1.0p-81 */
62 C6 = -1.1470297442401303e-11,		/* -0x19393412bd1529.0p-89 */
63 C7 =  4.7383039476436467e-14;		/*  0x1aac9d9af5c43e.0p-97 */
64 
65 long double
66 __kernel_cosl(long double x, long double y)
67 {
68 	long double hz,z,r,w;
69 
70 	z  = x*x;
71 	r  = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7))))));
72 	hz = 0.5*z;
73 	w  = one-hz;
74 	return w + (((one-w)-hz) + (z*r-x*y));
75 }
76