math: fix pow signed shift ub
[oweals/musl.git] / src / math / jnf.c
1 /* origin: FreeBSD /usr/src/lib/msun/src/e_jnf.c */
2 /*
3  * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.com.
4  */
5 /*
6  * ====================================================
7  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
8  *
9  * Developed at SunPro, a Sun Microsystems, Inc. business.
10  * Permission to use, copy, modify, and distribute this
11  * software is freely granted, provided that this notice
12  * is preserved.
13  * ====================================================
14  */
15
16 #define _GNU_SOURCE
17 #include "libm.h"
18
19 float jnf(int n, float x)
20 {
21         uint32_t ix;
22         int nm1, sign, i;
23         float a, b, temp;
24
25         GET_FLOAT_WORD(ix, x);
26         sign = ix>>31;
27         ix &= 0x7fffffff;
28         if (ix > 0x7f800000) /* nan */
29                 return x;
30
31         /* J(-n,x) = J(n,-x), use |n|-1 to avoid overflow in -n */
32         if (n == 0)
33                 return j0f(x);
34         if (n < 0) {
35                 nm1 = -(n+1);
36                 x = -x;
37                 sign ^= 1;
38         } else
39                 nm1 = n-1;
40         if (nm1 == 0)
41                 return j1f(x);
42
43         sign &= n;  /* even n: 0, odd n: signbit(x) */
44         x = fabsf(x);
45         if (ix == 0 || ix == 0x7f800000)  /* if x is 0 or inf */
46                 b = 0.0f;
47         else if (nm1 < x) {
48                 /* Safe to use J(n+1,x)=2n/x *J(n,x)-J(n-1,x) */
49                 a = j0f(x);
50                 b = j1f(x);
51                 for (i=0; i<nm1; ){
52                         i++;
53                         temp = b;
54                         b = b*(2.0f*i/x) - a;
55                         a = temp;
56                 }
57         } else {
58                 if (ix < 0x35800000) { /* x < 2**-20 */
59                         /* x is tiny, return the first Taylor expansion of J(n,x)
60                          * J(n,x) = 1/n!*(x/2)^n  - ...
61                          */
62                         if (nm1 > 8)  /* underflow */
63                                 nm1 = 8;
64                         temp = 0.5f * x;
65                         b = temp;
66                         a = 1.0f;
67                         for (i=2; i<=nm1+1; i++) {
68                                 a *= (float)i;    /* a = n! */
69                                 b *= temp;        /* b = (x/2)^n */
70                         }
71                         b = b/a;
72                 } else {
73                         /* use backward recurrence */
74                         /*                      x      x^2      x^2
75                          *  J(n,x)/J(n-1,x) =  ----   ------   ------   .....
76                          *                      2n  - 2(n+1) - 2(n+2)
77                          *
78                          *                      1      1        1
79                          *  (for large x)   =  ----  ------   ------   .....
80                          *                      2n   2(n+1)   2(n+2)
81                          *                      -- - ------ - ------ -
82                          *                       x     x         x
83                          *
84                          * Let w = 2n/x and h=2/x, then the above quotient
85                          * is equal to the continued fraction:
86                          *                  1
87                          *      = -----------------------
88                          *                     1
89                          *         w - -----------------
90                          *                        1
91                          *              w+h - ---------
92                          *                     w+2h - ...
93                          *
94                          * To determine how many terms needed, let
95                          * Q(0) = w, Q(1) = w(w+h) - 1,
96                          * Q(k) = (w+k*h)*Q(k-1) - Q(k-2),
97                          * When Q(k) > 1e4      good for single
98                          * When Q(k) > 1e9      good for double
99                          * When Q(k) > 1e17     good for quadruple
100                          */
101                         /* determine k */
102                         float t,q0,q1,w,h,z,tmp,nf;
103                         int k;
104
105                         nf = nm1+1.0f;
106                         w = 2*nf/x;
107                         h = 2/x;
108                         z = w+h;
109                         q0 = w;
110                         q1 = w*z - 1.0f;
111                         k = 1;
112                         while (q1 < 1.0e4f) {
113                                 k += 1;
114                                 z += h;
115                                 tmp = z*q1 - q0;
116                                 q0 = q1;
117                                 q1 = tmp;
118                         }
119                         for (t=0.0f, i=k; i>=0; i--)
120                                 t = 1.0f/(2*(i+nf)/x-t);
121                         a = t;
122                         b = 1.0f;
123                         /*  estimate log((2/x)^n*n!) = n*log(2/x)+n*ln(n)
124                          *  Hence, if n*(log(2n/x)) > ...
125                          *  single 8.8722839355e+01
126                          *  double 7.09782712893383973096e+02
127                          *  long double 1.1356523406294143949491931077970765006170e+04
128                          *  then recurrent value may overflow and the result is
129                          *  likely underflow to zero
130                          */
131                         tmp = nf*logf(fabsf(w));
132                         if (tmp < 88.721679688f) {
133                                 for (i=nm1; i>0; i--) {
134                                         temp = b;
135                                         b = 2.0f*i*b/x - a;
136                                         a = temp;
137                                 }
138                         } else {
139                                 for (i=nm1; i>0; i--){
140                                         temp = b;
141                                         b = 2.0f*i*b/x - a;
142                                         a = temp;
143                                         /* scale b to avoid spurious overflow */
144                                         if (b > 0x1p60f) {
145                                                 a /= b;
146                                                 t /= b;
147                                                 b = 1.0f;
148                                         }
149                                 }
150                         }
151                         z = j0f(x);
152                         w = j1f(x);
153                         if (fabsf(z) >= fabsf(w))
154                                 b = t*z/b;
155                         else
156                                 b = t*w/a;
157                 }
158         }
159         return sign ? -b : b;
160 }
161
162 float ynf(int n, float x)
163 {
164         uint32_t ix, ib;
165         int nm1, sign, i;
166         float a, b, temp;
167
168         GET_FLOAT_WORD(ix, x);
169         sign = ix>>31;
170         ix &= 0x7fffffff;
171         if (ix > 0x7f800000) /* nan */
172                 return x;
173         if (sign && ix != 0) /* x < 0 */
174                 return 0/0.0f;
175         if (ix == 0x7f800000)
176                 return 0.0f;
177
178         if (n == 0)
179                 return y0f(x);
180         if (n < 0) {
181                 nm1 = -(n+1);
182                 sign = n&1;
183         } else {
184                 nm1 = n-1;
185                 sign = 0;
186         }
187         if (nm1 == 0)
188                 return sign ? -y1f(x) : y1f(x);
189
190         a = y0f(x);
191         b = y1f(x);
192         /* quit if b is -inf */
193         GET_FLOAT_WORD(ib,b);
194         for (i = 0; i < nm1 && ib != 0xff800000; ) {
195                 i++;
196                 temp = b;
197                 b = (2.0f*i/x)*b - a;
198                 GET_FLOAT_WORD(ib, b);
199                 a = temp;
200         }
201         return sign ? -b : b;
202 }