xref: /freebsd/lib/msun/src/s_cbrtl.c (revision 9a14aa017b21c292740c00ee098195cd46642730)
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 #include <ieeefp.h>
22 
23 #include "fpmath.h"
24 #include "math.h"
25 #include "math_private.h"
26 
27 #define	BIAS	(LDBL_MAX_EXP - 1)
28 
29 static const unsigned
30     B1 = 709958130;	/* B1 = (127-127.0/3-0.03306235651)*2**23 */
31 
32 long double
33 cbrtl(long double x)
34 {
35 	union IEEEl2bits u, v;
36 	long double r, s, t, w;
37 	double dr, dt, dx;
38 	float ft, fx;
39 	uint32_t hx;
40 	uint16_t expsign;
41 	int k;
42 
43 	u.e = x;
44 	expsign = u.xbits.expsign;
45 	k = expsign & 0x7fff;
46 
47 	/*
48 	 * If x = +-Inf, then cbrt(x) = +-Inf.
49 	 * If x = NaN, then cbrt(x) = NaN.
50 	 */
51 	if (k == BIAS + LDBL_MAX_EXP)
52 		return (x + x);
53 
54 #ifdef __i386__
55 	fp_prec_t oprec;
56 
57 	oprec = fpgetprec();
58 	if (oprec != FP_PE)
59 		fpsetprec(FP_PE);
60 #endif
61 
62 	if (k == 0) {
63 		/* If x = +-0, then cbrt(x) = +-0. */
64 		if ((u.bits.manh | u.bits.manl) == 0) {
65 #ifdef __i386__
66 			if (oprec != FP_PE)
67 				fpsetprec(oprec);
68 #endif
69 			return (x);
70 	    	}
71 		/* Adjust subnormal numbers. */
72 		u.e *= 0x1.0p514;
73 		k = u.bits.exp;
74 		k -= BIAS + 514;
75  	} else
76 		k -= BIAS;
77 	u.xbits.expsign = BIAS;
78 	v.e = 1;
79 
80 	x = u.e;
81 	switch (k % 3) {
82 	case 1:
83 	case -2:
84 		x = 2*x;
85 		k--;
86 		break;
87 	case 2:
88 	case -1:
89 		x = 4*x;
90 		k -= 2;
91 		break;
92 	}
93 	v.xbits.expsign = (expsign & 0x8000) | (BIAS + k / 3);
94 
95 	/*
96 	 * The following is the guts of s_cbrtf, with the handling of
97 	 * special values removed and extra care for accuracy not taken,
98 	 * but with most of the extra accuracy not discarded.
99 	 */
100 
101 	/* ~5-bit estimate: */
102 	fx = x;
103 	GET_FLOAT_WORD(hx, fx);
104 	SET_FLOAT_WORD(ft, ((hx & 0x7fffffff) / 3 + B1));
105 
106 	/* ~16-bit estimate: */
107 	dx = x;
108 	dt = ft;
109 	dr = dt * dt * dt;
110 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
111 
112 	/* ~47-bit estimate: */
113 	dr = dt * dt * dt;
114 	dt = dt * (dx + dx + dr) / (dx + dr + dr);
115 
116 #if LDBL_MANT_DIG == 64
117 	/*
118 	 * dt is cbrtl(x) to ~47 bits (after x has been reduced to 1 <= x < 8).
119 	 * Round it away from zero to 32 bits (32 so that t*t is exact, and
120 	 * away from zero for technical reasons).
121 	 */
122 	volatile double vd2 = 0x1.0p32;
123 	volatile double vd1 = 0x1.0p-31;
124 	#define vd ((long double)vd2 + vd1)
125 
126 	t = dt + vd - 0x1.0p32;
127 #elif LDBL_MANT_DIG == 113
128 	/*
129 	 * Round dt away from zero to 47 bits.  Since we don't trust the 47,
130 	 * add 2 47-bit ulps instead of 1 to round up.  Rounding is slow and
131 	 * might be avoidable in this case, since on most machines dt will
132 	 * have been evaluated in 53-bit precision and the technical reasons
133 	 * for rounding up might not apply to either case in cbrtl() since
134 	 * dt is much more accurate than needed.
135 	 */
136 	t = dt + 0x2.0p-46 + 0x1.0p60L - 0x1.0p60;
137 #else
138 #error "Unsupported long double format"
139 #endif
140 
141 	/*
142      	 * Final step Newton iteration to 64 or 113 bits with
143 	 * error < 0.667 ulps
144 	 */
145 	s=t*t;				/* t*t is exact */
146 	r=x/s;				/* error <= 0.5 ulps; |r| < |t| */
147 	w=t+t;				/* t+t is exact */
148 	r=(r-t)/(w+r);			/* r-t is exact; w+r ~= 3*t */
149 	t=t+t*r;			/* error <= 0.5 + 0.5/3 + epsilon */
150 
151 	t *= v.e;
152 #ifdef __i386__
153 	if (oprec != FP_PE)
154 		fpsetprec(oprec);
155 #endif
156 	return (t);
157 }
158