| 1 | /* e_hypotl.c -- long double version of e_hypot.c. |
| 2 | * Conversion to long double by Ulrich Drepper, |
| 3 | * Cygnus Support, drepper@cygnus.com. |
| 4 | */ |
| 5 | |
| 6 | /* |
| 7 | * ==================================================== |
| 8 | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
| 9 | * |
| 10 | * Developed at SunPro, a Sun Microsystems, Inc. business. |
| 11 | * Permission to use, copy, modify, and distribute this |
| 12 | * software is freely granted, provided that this notice |
| 13 | * is preserved. |
| 14 | * ==================================================== |
| 15 | */ |
| 16 | |
| 17 | /* __ieee754_hypotl(x,y) |
| 18 | * |
| 19 | * Method : |
| 20 | * If (assume round-to-nearest) z=x*x+y*y |
| 21 | * has error less than sqrt(2)/2 ulp, than |
| 22 | * sqrt(z) has error less than 1 ulp (exercise). |
| 23 | * |
| 24 | * So, compute sqrt(x*x+y*y) with some care as |
| 25 | * follows to get the error below 1 ulp: |
| 26 | * |
| 27 | * Assume x>y>0; |
| 28 | * (if possible, set rounding to round-to-nearest) |
| 29 | * 1. if x > 2y use |
| 30 | * x1*x1+(y*y+(x2*(x+x1))) for x*x+y*y |
| 31 | * where x1 = x with lower 32 bits cleared, x2 = x-x1; else |
| 32 | * 2. if x <= 2y use |
| 33 | * t1*y1+((x-y)*(x-y)+(t1*y2+t2*y)) |
| 34 | * where t1 = 2x with lower 32 bits cleared, t2 = 2x-t1, |
| 35 | * y1= y with lower 32 bits chopped, y2 = y-y1. |
| 36 | * |
| 37 | * NOTE: scaling may be necessary if some argument is too |
| 38 | * large or too tiny |
| 39 | * |
| 40 | * Special cases: |
| 41 | * hypot(x,y) is INF if x or y is +INF or -INF; else |
| 42 | * hypot(x,y) is NAN if x or y is NAN. |
| 43 | * |
| 44 | * Accuracy: |
| 45 | * hypot(x,y) returns sqrt(x^2+y^2) with error less |
| 46 | * than 1 ulps (units in the last place) |
| 47 | */ |
| 48 | |
| 49 | #include <math.h> |
| 50 | #include <math_private.h> |
| 51 | #include <math-underflow.h> |
| 52 | #include <libm-alias-finite.h> |
| 53 | |
| 54 | long double __ieee754_hypotl(long double x, long double y) |
| 55 | { |
| 56 | long double a,b,t1,t2,y1,y2,w; |
| 57 | uint32_t j,k,ea,eb; |
| 58 | |
| 59 | GET_LDOUBLE_EXP(ea,x); |
| 60 | ea &= 0x7fff; |
| 61 | GET_LDOUBLE_EXP(eb,y); |
| 62 | eb &= 0x7fff; |
| 63 | if(eb > ea) {a=y;b=x;j=ea; ea=eb;eb=j;} else {a=x;b=y;} |
| 64 | SET_LDOUBLE_EXP(a,ea); /* a <- |a| */ |
| 65 | SET_LDOUBLE_EXP(b,eb); /* b <- |b| */ |
| 66 | if((ea-eb)>0x46) {return a+b;} /* x/y > 2**70 */ |
| 67 | k=0; |
| 68 | if(__builtin_expect(ea > 0x5f3f,0)) { /* a>2**8000 */ |
| 69 | if(ea == 0x7fff) { /* Inf or NaN */ |
| 70 | uint32_t exp __attribute__ ((unused)); |
| 71 | uint32_t high,low; |
| 72 | w = a+b; /* for sNaN */ |
| 73 | if (issignaling (a) || issignaling (b)) |
| 74 | return w; |
| 75 | GET_LDOUBLE_WORDS(exp,high,low,a); |
| 76 | if(((high&0x7fffffff)|low)==0) w = a; |
| 77 | GET_LDOUBLE_WORDS(exp,high,low,b); |
| 78 | if(((eb^0x7fff)|(high&0x7fffffff)|low)==0) w = b; |
| 79 | return w; |
| 80 | } |
| 81 | /* scale a and b by 2**-9600 */ |
| 82 | ea -= 0x2580; eb -= 0x2580; k += 9600; |
| 83 | SET_LDOUBLE_EXP(a,ea); |
| 84 | SET_LDOUBLE_EXP(b,eb); |
| 85 | } |
| 86 | if(__builtin_expect(eb < 0x20bf, 0)) { /* b < 2**-8000 */ |
| 87 | if(eb == 0) { /* subnormal b or 0 */ |
| 88 | uint32_t exp __attribute__ ((unused)); |
| 89 | uint32_t high,low; |
| 90 | GET_LDOUBLE_WORDS(exp,high,low,b); |
| 91 | if((high|low)==0) return a; |
| 92 | SET_LDOUBLE_WORDS(t1, 0x7ffd, 0x80000000, 0); /* t1=2^16382 */ |
| 93 | b *= t1; |
| 94 | a *= t1; |
| 95 | k -= 16382; |
| 96 | GET_LDOUBLE_EXP (ea, a); |
| 97 | GET_LDOUBLE_EXP (eb, b); |
| 98 | if (eb > ea) |
| 99 | { |
| 100 | t1 = a; |
| 101 | a = b; |
| 102 | b = t1; |
| 103 | j = ea; |
| 104 | ea = eb; |
| 105 | eb = j; |
| 106 | } |
| 107 | } else { /* scale a and b by 2^9600 */ |
| 108 | ea += 0x2580; /* a *= 2^9600 */ |
| 109 | eb += 0x2580; /* b *= 2^9600 */ |
| 110 | k -= 9600; |
| 111 | SET_LDOUBLE_EXP(a,ea); |
| 112 | SET_LDOUBLE_EXP(b,eb); |
| 113 | } |
| 114 | } |
| 115 | /* medium size a and b */ |
| 116 | w = a-b; |
| 117 | if (w>b) { |
| 118 | uint32_t high; |
| 119 | GET_LDOUBLE_MSW(high,a); |
| 120 | SET_LDOUBLE_WORDS(t1,ea,high,0); |
| 121 | t2 = a-t1; |
| 122 | w = sqrtl(t1*t1-(b*(-b)-t2*(a+t1))); |
| 123 | } else { |
| 124 | uint32_t high; |
| 125 | GET_LDOUBLE_MSW(high,b); |
| 126 | a = a+a; |
| 127 | SET_LDOUBLE_WORDS(y1,eb,high,0); |
| 128 | y2 = b - y1; |
| 129 | GET_LDOUBLE_MSW(high,a); |
| 130 | SET_LDOUBLE_WORDS(t1,ea+1,high,0); |
| 131 | t2 = a - t1; |
| 132 | w = sqrtl(t1*y1-(w*(-w)-(t1*y2+t2*b))); |
| 133 | } |
| 134 | if(k!=0) { |
| 135 | uint32_t exp; |
| 136 | t1 = 1.0; |
| 137 | GET_LDOUBLE_EXP(exp,t1); |
| 138 | SET_LDOUBLE_EXP(t1,exp+k); |
| 139 | w *= t1; |
| 140 | math_check_force_underflow_nonneg (w); |
| 141 | return w; |
| 142 | } else return w; |
| 143 | } |
| 144 | libm_alias_finite (__ieee754_hypotl, __hypotl) |
| 145 | |