1 
2 /* @(#)k_cos.c 5.1 93/09/24 */
3 /*
4  * ====================================================
5  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
6  *
7  * Developed at SunPro, 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 /*
15  * __kernel_cos( x,  y )
16  * kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164
17  * Input x is assumed to be bounded by ~pi/4 in magnitude.
18  * Input y is the tail of x.
19  *
20  * Algorithm
21  *	1. Since cos(-x) = cos(x), we need only to consider positive x.
22  *	2. if x < 2^-27 (hx<0x3e400000 0), return 1 with inexact if x!=0.
23  *	3. cos(x) is approximated by a polynomial of degree 14 on
24  *	   [0,pi/4]
25  *		  	                 4            14
26  *	   	cos(x) ~ 1 - x*x/2 + C1*x + ... + C6*x
27  *	   where the remez error is
28  *
29  * 	|              2     4     6     8     10    12     14 |     -58
30  * 	|cos(x)-(1-.5*x +C1*x +C2*x +C3*x +C4*x +C5*x  +C6*x  )| <= 2
31  * 	|    					               |
32  *
33  * 	               4     6     8     10    12     14
34  *	4. let r = C1*x +C2*x +C3*x +C4*x +C5*x  +C6*x  , then
35  *	       cos(x) = 1 - x*x/2 + r
36  *	   since cos(x+y) ~ cos(x) - sin(x)*y
37  *			  ~ cos(x) - x*y,
38  *	   a correction term is necessary in cos(x) and hence
39  *		cos(x+y) = 1 - (x*x/2 - (r - x*y))
40  *	   For better accuracy when x > 0.3, let qx = |x|/4 with
41  *	   the last 32 bits mask off, and if x > 0.78125, let qx = 0.28125.
42  *	   Then
43  *		cos(x+y) = (1-qx) - ((x*x/2-qx) - (r-x*y)).
44  *	   Note that 1-qx and (x*x/2-qx) is EXACT here, and the
45  *	   magnitude of the latter is at least a quarter of x*x/2,
46  *	   thus, reducing the rounding error in the subtraction.
47  */
48 
49 #include "fdlibm.h"
50 
51 #ifdef _NEED_FLOAT64
52 
53 static const __float64
54     one = _F_64(1.00000000000000000000e+00), /* 0x3FF00000, 0x00000000 */
55     C1 = _F_64(4.16666666666666019037e-02), /* 0x3FA55555, 0x5555554C */
56     C2 = _F_64(-1.38888888888741095749e-03), /* 0xBF56C16C, 0x16C15177 */
57     C3 = _F_64(2.48015872894767294178e-05), /* 0x3EFA01A0, 0x19CB1590 */
58     C4 = _F_64(-2.75573143513906633035e-07), /* 0xBE927E4F, 0x809C52AD */
59     C5 = _F_64(2.08757232129817482790e-09), /* 0x3E21EE9E, 0xBDB4B1C4 */
60     C6 = _F_64(-1.13596475577881948265e-11); /* 0xBDA8FAE9, 0xBE8838D4 */
61 
62 __float64
__kernel_cos(__float64 x,__float64 y)63 __kernel_cos(__float64 x, __float64 y)
64 {
65     __float64 a, hz, z, r, qx;
66     __int32_t ix;
67     GET_HIGH_WORD(ix, x);
68     ix &= 0x7fffffff; /* ix = |x|'s high word*/
69     if (ix < 0x3e400000) /* if x < 2**27 */
70         return one;
71     z = x * x;
72     r = z * (C1 + z * (C2 + z * (C3 + z * (C4 + z * (C5 + z * C6)))));
73     if (ix < 0x3FD33333) /* if |x| < 0.3 */
74         return one - (_F_64(0.5) * z - (z * r - x * y));
75     else {
76         if (ix > 0x3fe90000) { /* x > 0.78125 */
77             qx = _F_64(0.28125);
78         } else {
79             INSERT_WORDS(qx, ix - 0x00200000, 0); /* x/4 */
80         }
81         hz = _F_64(0.5) * z - qx;
82         a = one - qx;
83         return a - (hz - (z * r - x * y));
84     }
85 }
86 
87 #endif /* _NEED_FLOAT64 */
88