1176357Sdas/* From: @(#)k_cos.c 1.3 95/01/18 */
2176357Sdas/*
3176357Sdas * ====================================================
4176357Sdas * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5176357Sdas * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
6176357Sdas *
7176357Sdas * Developed at SunSoft, a Sun Microsystems, Inc. business.
8176357Sdas * Permission to use, copy, modify, and distribute this
9176357Sdas * software is freely granted, provided that this notice
10176357Sdas * is preserved.
11176357Sdas * ====================================================
12176357Sdas */
13176357Sdas
14176357Sdas#include <sys/cdefs.h>
15176357Sdas__FBSDID("$FreeBSD$");
16176357Sdas
17176357Sdas/*
18176357Sdas * ld80 version of k_cos.c.  See ../src/k_cos.c for most comments.
19176357Sdas */
20176357Sdas
21176357Sdas#include "math_private.h"
22176357Sdas
23176357Sdas/*
24176357Sdas * Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]:
25176357Sdas * |cos(x) - c(x)| < 2**-75.1
26176357Sdas *
27176357Sdas * The coefficients of c(x) were generated by a pari-gp script using
28176357Sdas * a Remez algorithm that searches for the best higher coefficients
29176357Sdas * after rounding leading coefficients to a specified precision.
30176357Sdas *
31176357Sdas * Simpler methods like Chebyshev or basic Remez barely suffice for
32176357Sdas * cos() in 64-bit precision, because we want the coefficient of x^2
33176357Sdas * to be precisely -0.5 so that multiplying by it is exact, and plain
34176357Sdas * rounding of the coefficients of a good polynomial approximation only
35176357Sdas * gives this up to about 64-bit precision.  Plain rounding also gives
36176357Sdas * a mediocre approximation for the coefficient of x^4, but a rounding
37176357Sdas * error of 0.5 ulps for this coefficient would only contribute ~0.01
38176357Sdas * ulps to the final error, so this is unimportant.  Rounding errors in
39176357Sdas * higher coefficients are even less important.
40176357Sdas *
41176357Sdas * In fact, coefficients above the x^4 one only need to have 53-bit
42176357Sdas * precision, and this is more efficient.  We get this optimization
43176357Sdas * almost for free from the complications needed to search for the best
44176357Sdas * higher coefficients.
45176357Sdas */
46176357Sdasstatic const double
47176357Sdasone = 1.0;
48176357Sdas
49176357Sdas#if defined(__amd64__) || defined(__i386__)
50176357Sdas/* Long double constants are slow on these arches, and broken on i386. */
51176357Sdasstatic const volatile double
52176357SdasC1hi = 0.041666666666666664,		/*  0x15555555555555.0p-57 */
53176357SdasC1lo = 2.2598839032744733e-18;		/*  0x14d80000000000.0p-111 */
54176357Sdas#define	C1	((long double)C1hi + C1lo)
55176357Sdas#else
56176357Sdasstatic const long double
57176357SdasC1 =  0.0416666666666666666136L;	/*  0xaaaaaaaaaaaaaa9b.0p-68 */
58176357Sdas#endif
59176357Sdas
60176357Sdasstatic const double
61176357SdasC2 = -0.0013888888888888874,		/* -0x16c16c16c16c10.0p-62 */
62176357SdasC3 =  0.000024801587301571716,		/*  0x1a01a01a018e22.0p-68 */
63176357SdasC4 = -0.00000027557319215507120,	/* -0x127e4fb7602f22.0p-74 */
64176357SdasC5 =  0.0000000020876754400407278,	/*  0x11eed8caaeccf1.0p-81 */
65176357SdasC6 = -1.1470297442401303e-11,		/* -0x19393412bd1529.0p-89 */
66176357SdasC7 =  4.7383039476436467e-14;		/*  0x1aac9d9af5c43e.0p-97 */
67176357Sdas
68176357Sdaslong double
69176357Sdas__kernel_cosl(long double x, long double y)
70176357Sdas{
71176357Sdas	long double hz,z,r,w;
72176357Sdas
73176357Sdas	z  = x*x;
74176357Sdas	r  = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7))))));
75176357Sdas	hz = 0.5*z;
76176357Sdas	w  = one-hz;
77176357Sdas	return w + (((one-w)-hz) + (z*r-x*y));
78176357Sdas}
79