| 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 | } |