1 | /* e_hypotl.c -- long double version of e_hypot.c. |
2 | * Conversion to long double by Jakub Jelinek, jakub@redhat.com. |
3 | */ |
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 | /* __ieee754_hypotl(x,y) |
17 | * |
18 | * Method : |
19 | * If (assume round-to-nearest) z=x*x+y*y |
20 | * has error less than sqrtl(2)/2 ulp, than |
21 | * sqrtl(z) has error less than 1 ulp (exercise). |
22 | * |
23 | * So, compute sqrtl(x*x+y*y) with some care as |
24 | * follows to get the error below 1 ulp: |
25 | * |
26 | * Assume x>y>0; |
27 | * (if possible, set rounding to round-to-nearest) |
28 | * 1. if x > 2y use |
29 | * x1*x1+(y*y+(x2*(x+x1))) for x*x+y*y |
30 | * where x1 = x with lower 64 bits cleared, x2 = x-x1; else |
31 | * 2. if x <= 2y use |
32 | * t1*y1+((x-y)*(x-y)+(t1*y2+t2*y)) |
33 | * where t1 = 2x with lower 64 bits cleared, t2 = 2x-t1, |
34 | * y1= y with lower 64 bits chopped, y2 = y-y1. |
35 | * |
36 | * NOTE: scaling may be necessary if some argument is too |
37 | * large or too tiny |
38 | * |
39 | * Special cases: |
40 | * hypotl(x,y) is INF if x or y is +INF or -INF; else |
41 | * hypotl(x,y) is NAN if x or y is NAN. |
42 | * |
43 | * Accuracy: |
44 | * hypotl(x,y) returns sqrtl(x^2+y^2) with error less |
45 | * than 1 ulps (units in the last place) |
46 | */ |
47 | |
48 | #include <math.h> |
49 | #include <math_private.h> |
50 | |
51 | _Float128 |
52 | __ieee754_hypotl(_Float128 x, _Float128 y) |
53 | { |
54 | _Float128 a,b,t1,t2,y1,y2,w; |
55 | int64_t j,k,ha,hb; |
56 | |
57 | GET_LDOUBLE_MSW64(ha,x); |
58 | ha &= 0x7fffffffffffffffLL; |
59 | GET_LDOUBLE_MSW64(hb,y); |
60 | hb &= 0x7fffffffffffffffLL; |
61 | if(hb > ha) {a=y;b=x;j=ha; ha=hb;hb=j;} else {a=x;b=y;} |
62 | SET_LDOUBLE_MSW64(a,ha); /* a <- |a| */ |
63 | SET_LDOUBLE_MSW64(b,hb); /* b <- |b| */ |
64 | if((ha-hb)>0x78000000000000LL) {return a+b;} /* x/y > 2**120 */ |
65 | k=0; |
66 | if(ha > 0x5f3f000000000000LL) { /* a>2**8000 */ |
67 | if(ha >= 0x7fff000000000000LL) { /* Inf or NaN */ |
68 | u_int64_t low; |
69 | w = a+b; /* for sNaN */ |
70 | if (issignaling (a) || issignaling (b)) |
71 | return w; |
72 | GET_LDOUBLE_LSW64(low,a); |
73 | if(((ha&0xffffffffffffLL)|low)==0) w = a; |
74 | GET_LDOUBLE_LSW64(low,b); |
75 | if(((hb^0x7fff000000000000LL)|low)==0) w = b; |
76 | return w; |
77 | } |
78 | /* scale a and b by 2**-9600 */ |
79 | ha -= 0x2580000000000000LL; |
80 | hb -= 0x2580000000000000LL; k += 9600; |
81 | SET_LDOUBLE_MSW64(a,ha); |
82 | SET_LDOUBLE_MSW64(b,hb); |
83 | } |
84 | if(hb < 0x20bf000000000000LL) { /* b < 2**-8000 */ |
85 | if(hb <= 0x0000ffffffffffffLL) { /* subnormal b or 0 */ |
86 | u_int64_t low; |
87 | GET_LDOUBLE_LSW64(low,b); |
88 | if((hb|low)==0) return a; |
89 | t1=0; |
90 | SET_LDOUBLE_MSW64(t1,0x7ffd000000000000LL); /* t1=2^16382 */ |
91 | b *= t1; |
92 | a *= t1; |
93 | k -= 16382; |
94 | GET_LDOUBLE_MSW64 (ha, a); |
95 | GET_LDOUBLE_MSW64 (hb, b); |
96 | if (hb > ha) |
97 | { |
98 | t1 = a; |
99 | a = b; |
100 | b = t1; |
101 | j = ha; |
102 | ha = hb; |
103 | hb = j; |
104 | } |
105 | } else { /* scale a and b by 2^9600 */ |
106 | ha += 0x2580000000000000LL; /* a *= 2^9600 */ |
107 | hb += 0x2580000000000000LL; /* b *= 2^9600 */ |
108 | k -= 9600; |
109 | SET_LDOUBLE_MSW64(a,ha); |
110 | SET_LDOUBLE_MSW64(b,hb); |
111 | } |
112 | } |
113 | /* medium size a and b */ |
114 | w = a-b; |
115 | if (w>b) { |
116 | t1 = 0; |
117 | SET_LDOUBLE_MSW64(t1,ha); |
118 | t2 = a-t1; |
119 | w = __ieee754_sqrtl(t1*t1-(b*(-b)-t2*(a+t1))); |
120 | } else { |
121 | a = a+a; |
122 | y1 = 0; |
123 | SET_LDOUBLE_MSW64(y1,hb); |
124 | y2 = b - y1; |
125 | t1 = 0; |
126 | SET_LDOUBLE_MSW64(t1,ha+0x0001000000000000LL); |
127 | t2 = a - t1; |
128 | w = __ieee754_sqrtl(t1*y1-(w*(-w)-(t1*y2+t2*b))); |
129 | } |
130 | if(k!=0) { |
131 | u_int64_t high; |
132 | t1 = 1; |
133 | GET_LDOUBLE_MSW64(high,t1); |
134 | SET_LDOUBLE_MSW64(t1,high+(k<<48)); |
135 | w *= t1; |
136 | math_check_force_underflow_nonneg (w); |
137 | return w; |
138 | } else return w; |
139 | } |
140 | strong_alias (__ieee754_hypotl, __hypotl_finite) |
141 | |