1 |
/* $MidnightBSD$ */ |
2 |
/* From: @(#)e_rem_pio2.c 1.4 95/01/18 */ |
3 |
/* |
4 |
* ==================================================== |
5 |
* Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
6 |
* Copyright (c) 2008 Steven G. Kargl, David Schultz, Bruce D. Evans. |
7 |
* |
8 |
* Developed at SunSoft, a Sun Microsystems, Inc. business. |
9 |
* Permission to use, copy, modify, and distribute this |
10 |
* software is freely granted, provided that this notice |
11 |
* is preserved. |
12 |
* ==================================================== |
13 |
* |
14 |
* Optimized by Bruce D. Evans. |
15 |
*/ |
16 |
|
17 |
#include <sys/cdefs.h> |
18 |
__FBSDID("$FreeBSD: stable/10/lib/msun/ld80/e_rem_pio2l.h 223262 2011-06-18 13:56:33Z benl $"); |
19 |
|
20 |
/* ld80 version of __ieee754_rem_pio2l(x,y) |
21 |
* |
22 |
* return the remainder of x rem pi/2 in y[0]+y[1] |
23 |
* use __kernel_rem_pio2() |
24 |
*/ |
25 |
|
26 |
#include <float.h> |
27 |
|
28 |
#include "math.h" |
29 |
#include "math_private.h" |
30 |
#include "fpmath.h" |
31 |
|
32 |
#define BIAS (LDBL_MAX_EXP - 1) |
33 |
|
34 |
/* |
35 |
* invpio2: 64 bits of 2/pi |
36 |
* pio2_1: first 39 bits of pi/2 |
37 |
* pio2_1t: pi/2 - pio2_1 |
38 |
* pio2_2: second 39 bits of pi/2 |
39 |
* pio2_2t: pi/2 - (pio2_1+pio2_2) |
40 |
* pio2_3: third 39 bits of pi/2 |
41 |
* pio2_3t: pi/2 - (pio2_1+pio2_2+pio2_3) |
42 |
*/ |
43 |
|
44 |
static const double |
45 |
zero = 0.00000000000000000000e+00, /* 0x00000000, 0x00000000 */ |
46 |
two24 = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */ |
47 |
pio2_1 = 1.57079632679597125389e+00, /* 0x3FF921FB, 0x54444000 */ |
48 |
pio2_2 = -1.07463465549783099519e-12, /* -0x12e7b967674000.0p-92 */ |
49 |
pio2_3 = 6.36831716351370313614e-25; /* 0x18a2e037074000.0p-133 */ |
50 |
|
51 |
#if defined(__amd64__) || defined(__i386__) |
52 |
/* Long double constants are slow on these arches, and broken on i386. */ |
53 |
static const volatile double |
54 |
invpio2hi = 6.3661977236758138e-01, /* 0x145f306dc9c883.0p-53 */ |
55 |
invpio2lo = -3.9356538861223811e-17, /* -0x16b00000000000.0p-107 */ |
56 |
pio2_1thi = -1.0746346554971943e-12, /* -0x12e7b9676733af.0p-92 */ |
57 |
pio2_1tlo = 8.8451028997905949e-29, /* 0x1c080000000000.0p-146 */ |
58 |
pio2_2thi = 6.3683171635109499e-25, /* 0x18a2e03707344a.0p-133 */ |
59 |
pio2_2tlo = 2.3183081793789774e-41, /* 0x10280000000000.0p-187 */ |
60 |
pio2_3thi = -2.7529965190440717e-37, /* -0x176b7ed8fbbacc.0p-174 */ |
61 |
pio2_3tlo = -4.2006647512740502e-54; /* -0x19c00000000000.0p-230 */ |
62 |
#define invpio2 ((long double)invpio2hi + invpio2lo) |
63 |
#define pio2_1t ((long double)pio2_1thi + pio2_1tlo) |
64 |
#define pio2_2t ((long double)pio2_2thi + pio2_2tlo) |
65 |
#define pio2_3t ((long double)pio2_3thi + pio2_3tlo) |
66 |
#else |
67 |
static const long double |
68 |
invpio2 = 6.36619772367581343076e-01L, /* 0xa2f9836e4e44152a.0p-64 */ |
69 |
pio2_1t = -1.07463465549719416346e-12L, /* -0x973dcb3b399d747f.0p-103 */ |
70 |
pio2_2t = 6.36831716351095013979e-25L, /* 0xc51701b839a25205.0p-144 */ |
71 |
pio2_3t = -2.75299651904407171810e-37L; /* -0xbb5bf6c7ddd660ce.0p-185 */ |
72 |
#endif |
73 |
|
74 |
static inline __always_inline int |
75 |
__ieee754_rem_pio2l(long double x, long double *y) |
76 |
{ |
77 |
union IEEEl2bits u,u1; |
78 |
long double z,w,t,r,fn; |
79 |
double tx[3],ty[2]; |
80 |
int e0,ex,i,j,nx,n; |
81 |
int16_t expsign; |
82 |
|
83 |
u.e = x; |
84 |
expsign = u.xbits.expsign; |
85 |
ex = expsign & 0x7fff; |
86 |
if (ex < BIAS + 25 || (ex == BIAS + 25 && u.bits.manh < 0xc90fdaa2)) { |
87 |
/* |x| ~< 2^25*(pi/2), medium size */ |
88 |
/* Use a specialized rint() to get fn. Assume round-to-nearest. */ |
89 |
fn = x*invpio2+0x1.8p63; |
90 |
fn = fn-0x1.8p63; |
91 |
#ifdef HAVE_EFFICIENT_IRINT |
92 |
n = irint(fn); |
93 |
#else |
94 |
n = fn; |
95 |
#endif |
96 |
r = x-fn*pio2_1; |
97 |
w = fn*pio2_1t; /* 1st round good to 102 bit */ |
98 |
{ |
99 |
union IEEEl2bits u2; |
100 |
int ex1; |
101 |
j = ex; |
102 |
y[0] = r-w; |
103 |
u2.e = y[0]; |
104 |
ex1 = u2.xbits.expsign & 0x7fff; |
105 |
i = j-ex1; |
106 |
if(i>22) { /* 2nd iteration needed, good to 141 */ |
107 |
t = r; |
108 |
w = fn*pio2_2; |
109 |
r = t-w; |
110 |
w = fn*pio2_2t-((t-r)-w); |
111 |
y[0] = r-w; |
112 |
u2.e = y[0]; |
113 |
ex1 = u2.xbits.expsign & 0x7fff; |
114 |
i = j-ex1; |
115 |
if(i>61) { /* 3rd iteration need, 180 bits acc */ |
116 |
t = r; /* will cover all possible cases */ |
117 |
w = fn*pio2_3; |
118 |
r = t-w; |
119 |
w = fn*pio2_3t-((t-r)-w); |
120 |
y[0] = r-w; |
121 |
} |
122 |
} |
123 |
} |
124 |
y[1] = (r-y[0])-w; |
125 |
return n; |
126 |
} |
127 |
/* |
128 |
* all other (large) arguments |
129 |
*/ |
130 |
if(ex==0x7fff) { /* x is inf or NaN */ |
131 |
y[0]=y[1]=x-x; return 0; |
132 |
} |
133 |
/* set z = scalbn(|x|,ilogb(x)-23) */ |
134 |
u1.e = x; |
135 |
e0 = ex - BIAS - 23; /* e0 = ilogb(|x|)-23; */ |
136 |
u1.xbits.expsign = ex - e0; |
137 |
z = u1.e; |
138 |
for(i=0;i<2;i++) { |
139 |
tx[i] = (double)((int32_t)(z)); |
140 |
z = (z-tx[i])*two24; |
141 |
} |
142 |
tx[2] = z; |
143 |
nx = 3; |
144 |
while(tx[nx-1]==zero) nx--; /* skip zero term */ |
145 |
n = __kernel_rem_pio2(tx,ty,e0,nx,2); |
146 |
r = (long double)ty[0] + ty[1]; |
147 |
w = ty[1] - (r - ty[0]); |
148 |
if(expsign<0) {y[0] = -r; y[1] = -w; return -n;} |
149 |
y[0] = r; y[1] = w; return n; |
150 |
} |