xref: /dflybsd-src/contrib/openbsd_libm/src/ld80/k_cosl.c (revision 4382f29d99a100bd77a81697c2f699c11f6a472a)
1*05a0b428SJohn Marino /*	$OpenBSD: k_cosl.c,v 1.1 2008/12/09 20:00:35 martynas Exp $	*/
2*05a0b428SJohn Marino /* From: @(#)k_cos.c 1.3 95/01/18 */
3*05a0b428SJohn Marino /*
4*05a0b428SJohn Marino  * ====================================================
5*05a0b428SJohn Marino  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6*05a0b428SJohn Marino  * Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans.
7*05a0b428SJohn Marino  *
8*05a0b428SJohn Marino  * Developed at SunSoft, a Sun Microsystems, Inc. business.
9*05a0b428SJohn Marino  * Permission to use, copy, modify, and distribute this
10*05a0b428SJohn Marino  * software is freely granted, provided that this notice
11*05a0b428SJohn Marino  * is preserved.
12*05a0b428SJohn Marino  * ====================================================
13*05a0b428SJohn Marino  */
14*05a0b428SJohn Marino 
15*05a0b428SJohn Marino /*
16*05a0b428SJohn Marino  * ld80 version of k_cos.c.  See ../k_cos.c for most comments.
17*05a0b428SJohn Marino  */
18*05a0b428SJohn Marino 
19*05a0b428SJohn Marino #include "math_private.h"
20*05a0b428SJohn Marino 
21*05a0b428SJohn Marino /*
22*05a0b428SJohn Marino  * Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]:
23*05a0b428SJohn Marino  * |cos(x) - c(x)| < 2**-75.1
24*05a0b428SJohn Marino  *
25*05a0b428SJohn Marino  * The coefficients of c(x) were generated by a pari-gp script using
26*05a0b428SJohn Marino  * a Remez algorithm that searches for the best higher coefficients
27*05a0b428SJohn Marino  * after rounding leading coefficients to a specified precision.
28*05a0b428SJohn Marino  *
29*05a0b428SJohn Marino  * Simpler methods like Chebyshev or basic Remez barely suffice for
30*05a0b428SJohn Marino  * cos() in 64-bit precision, because we want the coefficient of x^2
31*05a0b428SJohn Marino  * to be precisely -0.5 so that multiplying by it is exact, and plain
32*05a0b428SJohn Marino  * rounding of the coefficients of a good polynomial approximation only
33*05a0b428SJohn Marino  * gives this up to about 64-bit precision.  Plain rounding also gives
34*05a0b428SJohn Marino  * a mediocre approximation for the coefficient of x^4, but a rounding
35*05a0b428SJohn Marino  * error of 0.5 ulps for this coefficient would only contribute ~0.01
36*05a0b428SJohn Marino  * ulps to the final error, so this is unimportant.  Rounding errors in
37*05a0b428SJohn Marino  * higher coefficients are even less important.
38*05a0b428SJohn Marino  *
39*05a0b428SJohn Marino  * In fact, coefficients above the x^4 one only need to have 53-bit
40*05a0b428SJohn Marino  * precision, and this is more efficient.  We get this optimization
41*05a0b428SJohn Marino  * almost for free from the complications needed to search for the best
42*05a0b428SJohn Marino  * higher coefficients.
43*05a0b428SJohn Marino  */
44*05a0b428SJohn Marino static const double
45*05a0b428SJohn Marino one = 1.0;
46*05a0b428SJohn Marino 
47*05a0b428SJohn Marino #if defined(__amd64__) || defined(__i386__)
48*05a0b428SJohn Marino /* Long double constants are slow on these arches, and broken on i386. */
49*05a0b428SJohn Marino static const volatile double
50*05a0b428SJohn Marino C1hi = 0.041666666666666664,		/*  0x15555555555555.0p-57 */
51*05a0b428SJohn Marino C1lo = 2.2598839032744733e-18;		/*  0x14d80000000000.0p-111 */
52*05a0b428SJohn Marino #define	C1	((long double)C1hi + C1lo)
53*05a0b428SJohn Marino #else
54*05a0b428SJohn Marino static const long double
55*05a0b428SJohn Marino C1 =  0.0416666666666666666136L;	/*  0xaaaaaaaaaaaaaa9b.0p-68 */
56*05a0b428SJohn Marino #endif
57*05a0b428SJohn Marino 
58*05a0b428SJohn Marino static const double
59*05a0b428SJohn Marino C2 = -0.0013888888888888874,		/* -0x16c16c16c16c10.0p-62 */
60*05a0b428SJohn Marino C3 =  0.000024801587301571716,		/*  0x1a01a01a018e22.0p-68 */
61*05a0b428SJohn Marino C4 = -0.00000027557319215507120,	/* -0x127e4fb7602f22.0p-74 */
62*05a0b428SJohn Marino C5 =  0.0000000020876754400407278,	/*  0x11eed8caaeccf1.0p-81 */
63*05a0b428SJohn Marino C6 = -1.1470297442401303e-11,		/* -0x19393412bd1529.0p-89 */
64*05a0b428SJohn Marino C7 =  4.7383039476436467e-14;		/*  0x1aac9d9af5c43e.0p-97 */
65*05a0b428SJohn Marino 
66*05a0b428SJohn Marino long double
__kernel_cosl(long double x,long double y)67*05a0b428SJohn Marino __kernel_cosl(long double x, long double y)
68*05a0b428SJohn Marino {
69*05a0b428SJohn Marino 	long double hz,z,r,w;
70*05a0b428SJohn Marino 
71*05a0b428SJohn Marino 	z  = x*x;
72*05a0b428SJohn Marino 	r  = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7))))));
73*05a0b428SJohn Marino 	hz = 0.5*z;
74*05a0b428SJohn Marino 	w  = one-hz;
75*05a0b428SJohn Marino 	return w + (((one-w)-hz) + (z*r-x*y));
76*05a0b428SJohn Marino }
77