xref: /freebsd/lib/msun/ld80/k_cosl.c (revision 4543ef516683042d46f3bd3bb8a4f3f746e00499)
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 /*
14  * ld80 version of k_cos.c.  See ../src/k_cos.c for most comments.
15  */
16 
17 #include "math_private.h"
18 
19 /*
20  * Domain [-0.7854, 0.7854], range ~[-2.43e-23, 2.425e-23]:
21  * |cos(x) - c(x)| < 2**-75.1
22  *
23  * The coefficients of c(x) were generated by a pari-gp script using
24  * a Remez algorithm that searches for the best higher coefficients
25  * after rounding leading coefficients to a specified precision.
26  *
27  * Simpler methods like Chebyshev or basic Remez barely suffice for
28  * cos() in 64-bit precision, because we want the coefficient of x^2
29  * to be precisely -0.5 so that multiplying by it is exact, and plain
30  * rounding of the coefficients of a good polynomial approximation only
31  * gives this up to about 64-bit precision.  Plain rounding also gives
32  * a mediocre approximation for the coefficient of x^4, but a rounding
33  * error of 0.5 ulps for this coefficient would only contribute ~0.01
34  * ulps to the final error, so this is unimportant.  Rounding errors in
35  * higher coefficients are even less important.
36  *
37  * In fact, coefficients above the x^4 one only need to have 53-bit
38  * precision, and this is more efficient.  We get this optimization
39  * almost for free from the complications needed to search for the best
40  * higher coefficients.
41  */
42 static const double
43 one = 1.0;
44 
45 #if defined(__amd64__) || defined(__i386__)
46 /* Long double constants are slow on these arches, and broken on i386. */
47 static const volatile double
48 C1hi = 0.041666666666666664,		/*  0x15555555555555.0p-57 */
49 C1lo = 2.2598839032744733e-18;		/*  0x14d80000000000.0p-111 */
50 #define	C1	((long double)C1hi + C1lo)
51 #else
52 static const long double
53 C1 =  0.0416666666666666666136L;	/*  0xaaaaaaaaaaaaaa9b.0p-68 */
54 #endif
55 
56 static const double
57 C2 = -0.0013888888888888874,		/* -0x16c16c16c16c10.0p-62 */
58 C3 =  0.000024801587301571716,		/*  0x1a01a01a018e22.0p-68 */
59 C4 = -0.00000027557319215507120,	/* -0x127e4fb7602f22.0p-74 */
60 C5 =  0.0000000020876754400407278,	/*  0x11eed8caaeccf1.0p-81 */
61 C6 = -1.1470297442401303e-11,		/* -0x19393412bd1529.0p-89 */
62 C7 =  4.7383039476436467e-14;		/*  0x1aac9d9af5c43e.0p-97 */
63 
64 long double
65 __kernel_cosl(long double x, long double y)
66 {
67 	long double hz,z,r,w;
68 
69 	z  = x*x;
70 	r  = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*(C6+z*C7))))));
71 	hz = 0.5*z;
72 	w  = one-hz;
73 	return w + (((one-w)-hz) + (z*r-x*y));
74 }
75