e_rem_pio2f.c revision 151221
1/* e_rem_pio2f.c -- float version of e_rem_pio2.c 2 * Conversion to float by Ian Lance Taylor, Cygnus Support, ian@cygnus.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#ifndef lint 17static char rcsid[] = "$FreeBSD: head/lib/msun/src/e_rem_pio2f.c 151221 2005-10-10 20:02:02Z bde $"; 18#endif 19 20/* __ieee754_rem_pio2f(x,y) 21 * 22 * return the remainder of x rem pi/2 in y[0]+y[1] 23 * use __kernel_rem_pio2f() 24 */ 25 26#include "math.h" 27#include "math_private.h" 28 29/* 30 * Table of constants for 2/pi, 396 Hex digits (476 decimal) of 2/pi 31 */ 32static const int32_t two_over_pi[] = { 330xA2, 0xF9, 0x83, 0x6E, 0x4E, 0x44, 0x15, 0x29, 0xFC, 340x27, 0x57, 0xD1, 0xF5, 0x34, 0xDD, 0xC0, 0xDB, 0x62, 350x95, 0x99, 0x3C, 0x43, 0x90, 0x41, 0xFE, 0x51, 0x63, 360xAB, 0xDE, 0xBB, 0xC5, 0x61, 0xB7, 0x24, 0x6E, 0x3A, 370x42, 0x4D, 0xD2, 0xE0, 0x06, 0x49, 0x2E, 0xEA, 0x09, 380xD1, 0x92, 0x1C, 0xFE, 0x1D, 0xEB, 0x1C, 0xB1, 0x29, 390xA7, 0x3E, 0xE8, 0x82, 0x35, 0xF5, 0x2E, 0xBB, 0x44, 400x84, 0xE9, 0x9C, 0x70, 0x26, 0xB4, 0x5F, 0x7E, 0x41, 410x39, 0x91, 0xD6, 0x39, 0x83, 0x53, 0x39, 0xF4, 0x9C, 420x84, 0x5F, 0x8B, 0xBD, 0xF9, 0x28, 0x3B, 0x1F, 0xF8, 430x97, 0xFF, 0xDE, 0x05, 0x98, 0x0F, 0xEF, 0x2F, 0x11, 440x8B, 0x5A, 0x0A, 0x6D, 0x1F, 0x6D, 0x36, 0x7E, 0xCF, 450x27, 0xCB, 0x09, 0xB7, 0x4F, 0x46, 0x3F, 0x66, 0x9E, 460x5F, 0xEA, 0x2D, 0x75, 0x27, 0xBA, 0xC7, 0xEB, 0xE5, 470xF1, 0x7B, 0x3D, 0x07, 0x39, 0xF7, 0x8A, 0x52, 0x92, 480xEA, 0x6B, 0xFB, 0x5F, 0xB1, 0x1F, 0x8D, 0x5D, 0x08, 490x56, 0x03, 0x30, 0x46, 0xFC, 0x7B, 0x6B, 0xAB, 0xF0, 500xCF, 0xBC, 0x20, 0x9A, 0xF4, 0x36, 0x1D, 0xA9, 0xE3, 510x91, 0x61, 0x5E, 0xE6, 0x1B, 0x08, 0x65, 0x99, 0x85, 520x5F, 0x14, 0xA0, 0x68, 0x40, 0x8D, 0xFF, 0xD8, 0x80, 530x4D, 0x73, 0x27, 0x31, 0x06, 0x06, 0x15, 0x56, 0xCA, 540x73, 0xA8, 0xC9, 0x60, 0xE2, 0x7B, 0xC0, 0x8C, 0x6B, 55}; 56 57/* 58 * This array is like the one in e_rem_pio2.c, but the numbers are 59 * single precision and the last few bits (8 here) are ignored by 60 * masking them off in the float word instead of by omitting the low 61 * word. 62 * 63 * Masking off 8 bits is not enough, but we defer further masking to 64 * runtime so that the mask is easy to change. We now mask off 21 65 * bits, which is the smallest number that makes the "quick check no 66 * cancellation" detect all cancellations for cases that it is used. 67 * It doesn't detect all non-cancellations, especiallly for small 68 * multiples of pi/2, but then the non-quick code selects the best 69 * approximation of pi/2 to use. The result is that arg reduction is 70 * always done with between 8 or 9 and 17 bits of extra precision in 71 * the medium-multiple case. With only 8 bits masked of we had 72 * negative extra precision in some cases starting near +-13*pi/2. 73 */ 74static const int32_t npio2_hw[] = { 750x3fc90f00, 0x40490f00, 0x4096cb00, 0x40c90f00, 0x40fb5300, 0x4116cb00, 760x412fed00, 0x41490f00, 0x41623100, 0x417b5300, 0x418a3a00, 0x4196cb00, 770x41a35c00, 0x41afed00, 0x41bc7e00, 0x41c90f00, 0x41d5a000, 0x41e23100, 780x41eec200, 0x41fb5300, 0x4203f200, 0x420a3a00, 0x42108300, 0x4216cb00, 790x421d1400, 0x42235c00, 0x4229a500, 0x422fed00, 0x42363600, 0x423c7e00, 800x4242c700, 0x42490f00 81}; 82 83/* 84 * invpio2: 24 bits of 2/pi 85 * pio2_1: first 17 bit of pi/2 86 * pio2_1t: pi/2 - pio2_1 87 * pio2_2: second 17 bit of pi/2 88 * pio2_2t: pi/2 - (pio2_1+pio2_2) 89 * pio2_3: third 17 bit of pi/2 90 * pio2_3t: pi/2 - (pio2_1+pio2_2+pio2_3) 91 */ 92 93static const float 94zero = 0.0000000000e+00, /* 0x00000000 */ 95half = 5.0000000000e-01, /* 0x3f000000 */ 96two8 = 2.5600000000e+02, /* 0x43800000 */ 97invpio2 = 6.3661980629e-01, /* 0x3f22f984 */ 98pio2_1 = 1.5707855225e+00, /* 0x3fc90f80 */ 99pio2_1t = 1.0804334124e-05, /* 0x37354443 */ 100pio2_2 = 1.0804273188e-05, /* 0x37354400 */ 101pio2_2t = 6.0770999344e-11, /* 0x2e85a308 */ 102pio2_3 = 6.0770943833e-11, /* 0x2e85a300 */ 103pio2_3t = 6.1232342629e-17; /* 0x248d3132 */ 104 105 int32_t __ieee754_rem_pio2f(float x, float *y) 106{ 107 float z,w,t,r,fn; 108 float tx[3]; 109 int32_t e0,i,j,nx,n,ix,hx; 110 111 GET_FLOAT_WORD(hx,x); 112 ix = hx&0x7fffffff; 113 if(ix<=0x3f490fd8) /* |x| ~<= pi/4 , no need for reduction */ 114 {y[0] = x; y[1] = 0; return 0;} 115 if(ix<0x4016cbe4) { /* |x| < 3pi/4, special case with n=+-1 */ 116 if(hx>0) { 117 z = x - pio2_1; 118 if((ix&0xfffe0000)!=0x3fc80000) { /* 17+24 bit pi OK */ 119 y[0] = z - pio2_1t; 120 y[1] = (z-y[0])-pio2_1t; 121 } else { /* near pi/2, use 17+17+24 bit pi */ 122 z -= pio2_2; 123 y[0] = z - pio2_2t; 124 y[1] = (z-y[0])-pio2_2t; 125 } 126 return 1; 127 } else { /* negative x */ 128 z = x + pio2_1; 129 if((ix&0xfffe0000)!=0x3fc80000) { /* 17+24 bit pi OK */ 130 y[0] = z + pio2_1t; 131 y[1] = (z-y[0])+pio2_1t; 132 } else { /* near pi/2, use 17+17+24 bit pi */ 133 z += pio2_2; 134 y[0] = z + pio2_2t; 135 y[1] = (z-y[0])+pio2_2t; 136 } 137 return -1; 138 } 139 } 140 if(ix<=0x43490f80) { /* |x| ~<= 2^7*(pi/2), medium size */ 141 t = fabsf(x); 142 n = (int32_t) (t*invpio2+half); 143 fn = (float)n; 144 r = t-fn*pio2_1; 145 w = fn*pio2_1t; /* 1st round good to 40 bit */ 146 if(n<32&&(ix&0xffe00000)!=(npio2_hw[n-1]&0xffe00000)) { 147 y[0] = r-w; /* quick check no cancellation */ 148 } else { 149 u_int32_t high; 150 j = ix>>23; 151 y[0] = r-w; 152 GET_FLOAT_WORD(high,y[0]); 153 i = j-((high>>23)&0xff); 154 if(i>8) { /* 2nd iteration needed, good to 57 */ 155 t = r; 156 w = fn*pio2_2; 157 r = t-w; 158 w = fn*pio2_2t-((t-r)-w); 159 y[0] = r-w; 160 GET_FLOAT_WORD(high,y[0]); 161 i = j-((high>>23)&0xff); 162 if(i>25) { /* 3rd iteration need, 74 bits acc */ 163 t = r; /* will cover all possible cases */ 164 w = fn*pio2_3; 165 r = t-w; 166 w = fn*pio2_3t-((t-r)-w); 167 y[0] = r-w; 168 } 169 } 170 } 171 y[1] = (r-y[0])-w; 172 if(hx<0) {y[0] = -y[0]; y[1] = -y[1]; return -n;} 173 else return n; 174 } 175 /* 176 * all other (large) arguments 177 */ 178 if(ix>=0x7f800000) { /* x is inf or NaN */ 179 y[0]=y[1]=x-x; return 0; 180 } 181 /* set z = scalbn(|x|,ilogb(x)-7) */ 182 e0 = (ix>>23)-134; /* e0 = ilogb(z)-7; */ 183 SET_FLOAT_WORD(z, ix - ((int32_t)(e0<<23))); 184 for(i=0;i<2;i++) { 185 tx[i] = (float)((int32_t)(z)); 186 z = (z-tx[i])*two8; 187 } 188 tx[2] = z; 189 nx = 3; 190 while(tx[nx-1]==zero) nx--; /* skip zero term */ 191 n = __kernel_rem_pio2f(tx,y,e0,nx,2,two_over_pi); 192 if(hx<0) {y[0] = -y[0]; y[1] = -y[1]; return -n;} 193 return n; 194} 195