xref: /freebsd/lib/msun/src/s_cbrtl.c (revision f386f04f11679fd31731bce42208bb4363b79e75)
1 /*-
2  * ====================================================
3  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2009-2011, Bruce D. Evans, Steven G. Kargl, David Schultz.
5  *
6  * Developed at SunPro, 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  * The argument reduction and testing for exceptional cases was
13  * written by Steven G. Kargl with input from Bruce D. Evans
14  * and David A. Schultz.
15  */
16 
17 #include <sys/cdefs.h>
18 __FBSDID("$FreeBSD$");
19 
20 #include <float.h>
21 #ifdef __i386__
22 #include <ieeefp.h>
23 #endif
24 
25 #include "fpmath.h"
26 #include "math.h"
27 #include "math_private.h"
28 
29 #define	BIAS	(LDBL_MAX_EXP - 1)
30 
31 static const unsigned
32     B1 = 709958130;	/* B1 = (127-127.0/3-0.03306235651)*2**23 */
33 
34 long double
35 cbrtl(long double x)
36 {
37 	union IEEEl2bits u, v;
38 	long double r, s, t, w;
39 	double dr, dt, dx;
40 	float ft, fx;
41 	uint32_t hx;
42 	uint16_t expsign;
43 	int k;
44 
45 	u.e = x;
46 	expsign = u.xbits.expsign;
47 	k = expsign & 0x7fff;
48 
49 	/*
50 	 * If x = +-Inf, then cbrt(x) = +-Inf.
51 	 * If x = NaN, then cbrt(x) = NaN.
52 	 */
53 	if (k == BIAS + LDBL_MAX_EXP)
54 		return (x + x);
55 
56 	ENTERI();
57 	if (k == 0) {
58 		/* If x = +-0, then cbrt(x) = +-0. */
59 		if ((u.bits.manh | u.bits.manl) == 0)
60 			RETURNI(x);
61 		/* Adjust subnormal numbers. */
62 		u.e *= 0x1.0p514;
63 		k = u.bits.exp;
64 		k -= BIAS + 514;
65  	} else
66 		k -= BIAS;
67 	u.xbits.expsign = BIAS;
68 	v.e = 1;
69 
70 	x = u.e;
71 	switch (k % 3) {
72 	case 1:
73 	case -2:
74 		x = 2*x;
75 		k--;
76 		break;
77 	case 2:
78 	case -1:
79 		x = 4*x;
80 		k -= 2;
81 		break;
82 	}
83 	v.xbits.expsign = (expsign & 0x8000) | (BIAS + k / 3);
84 
85 	/*
86 	 * The following is the guts of s_cbrtf, with the handling of
87 	 * special values removed and extra care for accuracy not taken,
88 	 * but with most of the extra accuracy not discarded.
89 	 */
90 
91 	/* ~5-bit estimate: */
92 	fx = x;
93 	GET_FLOAT_WORD(hx, fx);
94 	SET_FLOAT_WORD(ft, ((hx & 0x7fffffff) / 3 + B1));
95 
96 	/* ~16-bit estimate: */
97 	dx = x;
98 	dt = ft;
99 	dr = dt * dt * dt;
100 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
101 
102 	/* ~47-bit estimate: */
103 	dr = dt * dt * dt;
104 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
105 
106 #if LDBL_MANT_DIG == 64
107 	/*
108 	 * dt is cbrtl(x) to ~47 bits (after x has been reduced to 1 <= x < 8).
109 	 * Round it away from zero to 32 bits (32 so that t*t is exact, and
110 	 * away from zero for technical reasons).
111 	 */
112 	volatile double vd2 = 0x1.0p32;
113 	volatile double vd1 = 0x1.0p-31;
114 	#define vd ((long double)vd2 + vd1)
115 
116 	t = dt + vd - 0x1.0p32;
117 #elif LDBL_MANT_DIG == 113
118 	/*
119 	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
120 	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
121 	 * might be avoidable in this case, since on most machines dt will
122 	 * have been evaluated in 53-bit precision and the technical reasons
123 	 * for rounding up might not apply to either case in cbrtl() since
124 	 * dt is much more accurate than needed.
125 	 */
126 	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
127 #else
128 #error "Unsupported long double format"
129 #endif
130 
131 	/*
132      	 * Final step Newton iteration to 64 or 113 bits with
133 	 * error < 0.667 ulps
134 	 */
135 	s=t*t;				/* t*t is exact */
136 	r=x/s;				/* error <= 0.5 ulps; |r| < |t| */
137 	w=t+t;				/* t+t is exact */
138 	r=(r-t)/(w+r);			/* r-t is exact; w+r ~= 3*t */
139 	t=t+t*r;			/* error <= 0.5 + 0.5/3 + epsilon */
140 
141 	t *= v.e;
142 	RETURNI(t);
143 }
144