0N/A/*
1879N/A * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
0N/A * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
0N/A *
0N/A * This code is free software; you can redistribute it and/or modify it
0N/A * under the terms of the GNU General Public License version 2 only, as
0N/A * published by the Free Software Foundation.
0N/A *
0N/A * This code is distributed in the hope that it will be useful, but WITHOUT
0N/A * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
0N/A * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
0N/A * version 2 for more details (a copy is included in the LICENSE file that
0N/A * accompanied this code).
0N/A *
0N/A * You should have received a copy of the GNU General Public License version
0N/A * 2 along with this work; if not, write to the Free Software Foundation,
0N/A * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
0N/A *
1472N/A * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1472N/A * or visit www.oracle.com if you need additional information or have any
1472N/A * questions.
0N/A *
0N/A */
0N/A
1879N/A#include "precompiled.hpp"
1879N/A#include "prims/jni.h"
1879N/A#include "runtime/interfaceSupport.hpp"
1879N/A#include "runtime/sharedRuntime.hpp"
0N/A
0N/A// This file contains copies of the fdlibm routines used by
0N/A// StrictMath. It turns out that it is almost always required to use
0N/A// these runtime routines; the Intel CPU doesn't meet the Java
0N/A// specification for sin/cos outside a certain limited argument range,
0N/A// and the SPARC CPU doesn't appear to have sin/cos instructions. It
0N/A// also turns out that avoiding the indirect call through function
0N/A// pointer out to libjava.so in SharedRuntime speeds these routines up
0N/A// by roughly 15% on both Win32/x86 and Solaris/SPARC.
0N/A
0N/A// Enabling optimizations in this file causes incorrect code to be
0N/A// generated; can not figure out how to turn down optimization for one
0N/A// file in the IDE on Windows
0N/A#ifdef WIN32
0N/A# pragma optimize ( "", off )
0N/A#endif
0N/A
1405N/A/* The above workaround now causes more problems with the latest MS compiler.
1405N/A * Visual Studio 2010's /GS option tries to guard against buffer overruns.
1405N/A * /GS is on by default if you specify optimizations, which we do globally
1405N/A * via /W3 /O2. However the above selective turning off of optimizations means
1405N/A * that /GS issues a warning "4748". And since we treat warnings as errors (/WX)
1405N/A * then the compilation fails. There are several possible solutions
1405N/A * (1) Remove that pragma above as obsolete with VS2010 - requires testing.
1405N/A * (2) Stop treating warnings as errors - would be a backward step
1405N/A * (3) Disable /GS - may help performance but you lose the security checks
1405N/A * (4) Disable the warning with "#pragma warning( disable : 4748 )"
1405N/A * (5) Disable planting the code with __declspec(safebuffers)
1405N/A * I've opted for (5) although we should investigate the local performance
1405N/A * benefits of (1) and global performance benefit of (3).
1405N/A */
1405N/A#if defined(WIN32) && (defined(_MSC_VER) && (_MSC_VER >= 1600))
1405N/A#define SAFEBUF __declspec(safebuffers)
1405N/A#else
1405N/A#define SAFEBUF
1405N/A#endif
1405N/A
0N/A#include <math.h>
0N/A
0N/A// VM_LITTLE_ENDIAN is #defined appropriately in the Makefiles
0N/A// [jk] this is not 100% correct because the float word order may different
0N/A// from the byte order (e.g. on ARM)
0N/A#ifdef VM_LITTLE_ENDIAN
0N/A# define __HI(x) *(1+(int*)&x)
0N/A# define __LO(x) *(int*)&x
0N/A#else
0N/A# define __HI(x) *(int*)&x
0N/A# define __LO(x) *(1+(int*)&x)
0N/A#endif
0N/A
0N/Astatic double copysignA(double x, double y) {
0N/A __HI(x) = (__HI(x)&0x7fffffff)|(__HI(y)&0x80000000);
0N/A return x;
0N/A}
0N/A
0N/A/*
0N/A * scalbn (double x, int n)
0N/A * scalbn(x,n) returns x* 2**n computed by exponent
0N/A * manipulation rather than by actually performing an
0N/A * exponentiation or a multiplication.
0N/A */
0N/A
0N/Astatic const double
0N/Atwo54 = 1.80143985094819840000e+16, /* 0x43500000, 0x00000000 */
0N/Atwom54 = 5.55111512312578270212e-17, /* 0x3C900000, 0x00000000 */
0N/AhugeX = 1.0e+300,
0N/Atiny = 1.0e-300;
0N/A
0N/Astatic double scalbnA (double x, int n) {
0N/A int k,hx,lx;
0N/A hx = __HI(x);
0N/A lx = __LO(x);
0N/A k = (hx&0x7ff00000)>>20; /* extract exponent */
0N/A if (k==0) { /* 0 or subnormal x */
0N/A if ((lx|(hx&0x7fffffff))==0) return x; /* +-0 */
0N/A x *= two54;
0N/A hx = __HI(x);
0N/A k = ((hx&0x7ff00000)>>20) - 54;
0N/A if (n< -50000) return tiny*x; /*underflow*/
0N/A }
0N/A if (k==0x7ff) return x+x; /* NaN or Inf */
0N/A k = k+n;
0N/A if (k > 0x7fe) return hugeX*copysignA(hugeX,x); /* overflow */
0N/A if (k > 0) /* normal result */
0N/A {__HI(x) = (hx&0x800fffff)|(k<<20); return x;}
0N/A if (k <= -54) {
0N/A if (n > 50000) /* in case integer overflow in n+k */
0N/A return hugeX*copysignA(hugeX,x); /*overflow*/
0N/A else return tiny*copysignA(tiny,x); /*underflow*/
0N/A }
0N/A k += 54; /* subnormal result */
0N/A __HI(x) = (hx&0x800fffff)|(k<<20);
0N/A return x*twom54;
0N/A}
0N/A
0N/A/*
0N/A * __kernel_rem_pio2(x,y,e0,nx,prec,ipio2)
0N/A * double x[],y[]; int e0,nx,prec; int ipio2[];
0N/A *
0N/A * __kernel_rem_pio2 return the last three digits of N with
0N/A * y = x - N*pi/2
0N/A * so that |y| < pi/2.
0N/A *
0N/A * The method is to compute the integer (mod 8) and fraction parts of
0N/A * (2/pi)*x without doing the full multiplication. In general we
0N/A * skip the part of the product that are known to be a huge integer (
0N/A * more accurately, = 0 mod 8 ). Thus the number of operations are
0N/A * independent of the exponent of the input.
0N/A *
0N/A * (2/pi) is represented by an array of 24-bit integers in ipio2[].
0N/A *
0N/A * Input parameters:
0N/A * x[] The input value (must be positive) is broken into nx
0N/A * pieces of 24-bit integers in double precision format.
0N/A * x[i] will be the i-th 24 bit of x. The scaled exponent
0N/A * of x[0] is given in input parameter e0 (i.e., x[0]*2^e0
0N/A * match x's up to 24 bits.
0N/A *
0N/A * Example of breaking a double positive z into x[0]+x[1]+x[2]:
0N/A * e0 = ilogb(z)-23
0N/A * z = scalbn(z,-e0)
0N/A * for i = 0,1,2
0N/A * x[i] = floor(z)
0N/A * z = (z-x[i])*2**24
0N/A *
0N/A *
0N/A * y[] ouput result in an array of double precision numbers.
0N/A * The dimension of y[] is:
0N/A * 24-bit precision 1
0N/A * 53-bit precision 2
0N/A * 64-bit precision 2
0N/A * 113-bit precision 3
0N/A * The actual value is the sum of them. Thus for 113-bit
0N/A * precsion, one may have to do something like:
0N/A *
0N/A * long double t,w,r_head, r_tail;
0N/A * t = (long double)y[2] + (long double)y[1];
0N/A * w = (long double)y[0];
0N/A * r_head = t+w;
0N/A * r_tail = w - (r_head - t);
0N/A *
0N/A * e0 The exponent of x[0]
0N/A *
0N/A * nx dimension of x[]
0N/A *
0N/A * prec an interger indicating the precision:
0N/A * 0 24 bits (single)
0N/A * 1 53 bits (double)
0N/A * 2 64 bits (extended)
0N/A * 3 113 bits (quad)
0N/A *
0N/A * ipio2[]
0N/A * integer array, contains the (24*i)-th to (24*i+23)-th
0N/A * bit of 2/pi after binary point. The corresponding
0N/A * floating value is
0N/A *
0N/A * ipio2[i] * 2^(-24(i+1)).
0N/A *
0N/A * External function:
0N/A * double scalbn(), floor();
0N/A *
0N/A *
0N/A * Here is the description of some local variables:
0N/A *
0N/A * jk jk+1 is the initial number of terms of ipio2[] needed
0N/A * in the computation. The recommended value is 2,3,4,
0N/A * 6 for single, double, extended,and quad.
0N/A *
0N/A * jz local integer variable indicating the number of
0N/A * terms of ipio2[] used.
0N/A *
0N/A * jx nx - 1
0N/A *
0N/A * jv index for pointing to the suitable ipio2[] for the
0N/A * computation. In general, we want
0N/A * ( 2^e0*x[0] * ipio2[jv-1]*2^(-24jv) )/8
0N/A * is an integer. Thus
0N/A * e0-3-24*jv >= 0 or (e0-3)/24 >= jv
0N/A * Hence jv = max(0,(e0-3)/24).
0N/A *
0N/A * jp jp+1 is the number of terms in PIo2[] needed, jp = jk.
0N/A *
0N/A * q[] double array with integral value, representing the
0N/A * 24-bits chunk of the product of x and 2/pi.
0N/A *
0N/A * q0 the corresponding exponent of q[0]. Note that the
0N/A * exponent for q[i] would be q0-24*i.
0N/A *
0N/A * PIo2[] double precision array, obtained by cutting pi/2
0N/A * into 24 bits chunks.
0N/A *
0N/A * f[] ipio2[] in floating point
0N/A *
0N/A * iq[] integer array by breaking up q[] in 24-bits chunk.
0N/A *
0N/A * fq[] final product of x*(2/pi) in fq[0],..,fq[jk]
0N/A *
0N/A * ih integer. If >0 it indicats q[] is >= 0.5, hence
0N/A * it also indicates the *sign* of the result.
0N/A *
0N/A */
0N/A
0N/A
0N/A/*
0N/A * Constants:
0N/A * The hexadecimal values are the intended ones for the following
0N/A * constants. The decimal values may be used, provided that the
0N/A * compiler will convert from decimal to binary accurately enough
0N/A * to produce the hexadecimal values shown.
0N/A */
0N/A
0N/A
0N/Astatic const int init_jk[] = {2,3,4,6}; /* initial value for jk */
0N/A
0N/Astatic const double PIo2[] = {
0N/A 1.57079625129699707031e+00, /* 0x3FF921FB, 0x40000000 */
0N/A 7.54978941586159635335e-08, /* 0x3E74442D, 0x00000000 */
0N/A 5.39030252995776476554e-15, /* 0x3CF84698, 0x80000000 */
0N/A 3.28200341580791294123e-22, /* 0x3B78CC51, 0x60000000 */
0N/A 1.27065575308067607349e-29, /* 0x39F01B83, 0x80000000 */
0N/A 1.22933308981111328932e-36, /* 0x387A2520, 0x40000000 */
0N/A 2.73370053816464559624e-44, /* 0x36E38222, 0x80000000 */
0N/A 2.16741683877804819444e-51, /* 0x3569F31D, 0x00000000 */
0N/A};
0N/A
0N/Astatic const double
0N/AzeroB = 0.0,
0N/Aone = 1.0,
0N/Atwo24B = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */
0N/Atwon24 = 5.96046447753906250000e-08; /* 0x3E700000, 0x00000000 */
0N/A
1405N/Astatic SAFEBUF int __kernel_rem_pio2(double *x, double *y, int e0, int nx, int prec, const int *ipio2) {
0N/A int jz,jx,jv,jp,jk,carry,n,iq[20],i,j,k,m,q0,ih;
0N/A double z,fw,f[20],fq[20],q[20];
0N/A
0N/A /* initialize jk*/
0N/A jk = init_jk[prec];
0N/A jp = jk;
0N/A
0N/A /* determine jx,jv,q0, note that 3>q0 */
0N/A jx = nx-1;
0N/A jv = (e0-3)/24; if(jv<0) jv=0;
0N/A q0 = e0-24*(jv+1);
0N/A
0N/A /* set up f[0] to f[jx+jk] where f[jx+jk] = ipio2[jv+jk] */
0N/A j = jv-jx; m = jx+jk;
0N/A for(i=0;i<=m;i++,j++) f[i] = (j<0)? zeroB : (double) ipio2[j];
0N/A
0N/A /* compute q[0],q[1],...q[jk] */
0N/A for (i=0;i<=jk;i++) {
0N/A for(j=0,fw=0.0;j<=jx;j++) fw += x[j]*f[jx+i-j]; q[i] = fw;
0N/A }
0N/A
0N/A jz = jk;
0N/Arecompute:
0N/A /* distill q[] into iq[] reversingly */
0N/A for(i=0,j=jz,z=q[jz];j>0;i++,j--) {
0N/A fw = (double)((int)(twon24* z));
0N/A iq[i] = (int)(z-two24B*fw);
0N/A z = q[j-1]+fw;
0N/A }
0N/A
0N/A /* compute n */
0N/A z = scalbnA(z,q0); /* actual value of z */
0N/A z -= 8.0*floor(z*0.125); /* trim off integer >= 8 */
0N/A n = (int) z;
0N/A z -= (double)n;
0N/A ih = 0;
0N/A if(q0>0) { /* need iq[jz-1] to determine n */
0N/A i = (iq[jz-1]>>(24-q0)); n += i;
0N/A iq[jz-1] -= i<<(24-q0);
0N/A ih = iq[jz-1]>>(23-q0);
0N/A }
0N/A else if(q0==0) ih = iq[jz-1]>>23;
0N/A else if(z>=0.5) ih=2;
0N/A
0N/A if(ih>0) { /* q > 0.5 */
0N/A n += 1; carry = 0;
0N/A for(i=0;i<jz ;i++) { /* compute 1-q */
0N/A j = iq[i];
0N/A if(carry==0) {
0N/A if(j!=0) {
0N/A carry = 1; iq[i] = 0x1000000- j;
0N/A }
0N/A } else iq[i] = 0xffffff - j;
0N/A }
0N/A if(q0>0) { /* rare case: chance is 1 in 12 */
0N/A switch(q0) {
0N/A case 1:
0N/A iq[jz-1] &= 0x7fffff; break;
0N/A case 2:
0N/A iq[jz-1] &= 0x3fffff; break;
0N/A }
0N/A }
0N/A if(ih==2) {
0N/A z = one - z;
0N/A if(carry!=0) z -= scalbnA(one,q0);
0N/A }
0N/A }
0N/A
0N/A /* check if recomputation is needed */
0N/A if(z==zeroB) {
0N/A j = 0;
0N/A for (i=jz-1;i>=jk;i--) j |= iq[i];
0N/A if(j==0) { /* need recomputation */
0N/A for(k=1;iq[jk-k]==0;k++); /* k = no. of terms needed */
0N/A
0N/A for(i=jz+1;i<=jz+k;i++) { /* add q[jz+1] to q[jz+k] */
0N/A f[jx+i] = (double) ipio2[jv+i];
0N/A for(j=0,fw=0.0;j<=jx;j++) fw += x[j]*f[jx+i-j];
0N/A q[i] = fw;
0N/A }
0N/A jz += k;
0N/A goto recompute;
0N/A }
0N/A }
0N/A
0N/A /* chop off zero terms */
0N/A if(z==0.0) {
0N/A jz -= 1; q0 -= 24;
0N/A while(iq[jz]==0) { jz--; q0-=24;}
0N/A } else { /* break z into 24-bit if neccessary */
0N/A z = scalbnA(z,-q0);
0N/A if(z>=two24B) {
0N/A fw = (double)((int)(twon24*z));
0N/A iq[jz] = (int)(z-two24B*fw);
0N/A jz += 1; q0 += 24;
0N/A iq[jz] = (int) fw;
0N/A } else iq[jz] = (int) z ;
0N/A }
0N/A
0N/A /* convert integer "bit" chunk to floating-point value */
0N/A fw = scalbnA(one,q0);
0N/A for(i=jz;i>=0;i--) {
0N/A q[i] = fw*(double)iq[i]; fw*=twon24;
0N/A }
0N/A
0N/A /* compute PIo2[0,...,jp]*q[jz,...,0] */
0N/A for(i=jz;i>=0;i--) {
0N/A for(fw=0.0,k=0;k<=jp&&k<=jz-i;k++) fw += PIo2[k]*q[i+k];
0N/A fq[jz-i] = fw;
0N/A }
0N/A
0N/A /* compress fq[] into y[] */
0N/A switch(prec) {
0N/A case 0:
0N/A fw = 0.0;
0N/A for (i=jz;i>=0;i--) fw += fq[i];
0N/A y[0] = (ih==0)? fw: -fw;
0N/A break;
0N/A case 1:
0N/A case 2:
0N/A fw = 0.0;
0N/A for (i=jz;i>=0;i--) fw += fq[i];
0N/A y[0] = (ih==0)? fw: -fw;
0N/A fw = fq[0]-fw;
0N/A for (i=1;i<=jz;i++) fw += fq[i];
0N/A y[1] = (ih==0)? fw: -fw;
0N/A break;
0N/A case 3: /* painful */
0N/A for (i=jz;i>0;i--) {
0N/A fw = fq[i-1]+fq[i];
0N/A fq[i] += fq[i-1]-fw;
0N/A fq[i-1] = fw;
0N/A }
0N/A for (i=jz;i>1;i--) {
0N/A fw = fq[i-1]+fq[i];
0N/A fq[i] += fq[i-1]-fw;
0N/A fq[i-1] = fw;
0N/A }
0N/A for (fw=0.0,i=jz;i>=2;i--) fw += fq[i];
0N/A if(ih==0) {
0N/A y[0] = fq[0]; y[1] = fq[1]; y[2] = fw;
0N/A } else {
0N/A y[0] = -fq[0]; y[1] = -fq[1]; y[2] = -fw;
0N/A }
0N/A }
0N/A return n&7;
0N/A}
0N/A
0N/A
0N/A/*
0N/A * ====================================================
1472N/A * Copyright (c) 1993 Oracle and/or its affilates. All rights reserved.
0N/A *
0N/A * Developed at SunPro, a Sun Microsystems, Inc. business.
0N/A * Permission to use, copy, modify, and distribute this
0N/A * software is freely granted, provided that this notice
0N/A * is preserved.
0N/A * ====================================================
0N/A *
0N/A */
0N/A
0N/A/* __ieee754_rem_pio2(x,y)
0N/A *
0N/A * return the remainder of x rem pi/2 in y[0]+y[1]
0N/A * use __kernel_rem_pio2()
0N/A */
0N/A
0N/A/*
0N/A * Table of constants for 2/pi, 396 Hex digits (476 decimal) of 2/pi
0N/A */
0N/Astatic const int two_over_pi[] = {
0N/A 0xA2F983, 0x6E4E44, 0x1529FC, 0x2757D1, 0xF534DD, 0xC0DB62,
0N/A 0x95993C, 0x439041, 0xFE5163, 0xABDEBB, 0xC561B7, 0x246E3A,
0N/A 0x424DD2, 0xE00649, 0x2EEA09, 0xD1921C, 0xFE1DEB, 0x1CB129,
0N/A 0xA73EE8, 0x8235F5, 0x2EBB44, 0x84E99C, 0x7026B4, 0x5F7E41,
0N/A 0x3991D6, 0x398353, 0x39F49C, 0x845F8B, 0xBDF928, 0x3B1FF8,
0N/A 0x97FFDE, 0x05980F, 0xEF2F11, 0x8B5A0A, 0x6D1F6D, 0x367ECF,
0N/A 0x27CB09, 0xB74F46, 0x3F669E, 0x5FEA2D, 0x7527BA, 0xC7EBE5,
0N/A 0xF17B3D, 0x0739F7, 0x8A5292, 0xEA6BFB, 0x5FB11F, 0x8D5D08,
0N/A 0x560330, 0x46FC7B, 0x6BABF0, 0xCFBC20, 0x9AF436, 0x1DA9E3,
0N/A 0x91615E, 0xE61B08, 0x659985, 0x5F14A0, 0x68408D, 0xFFD880,
0N/A 0x4D7327, 0x310606, 0x1556CA, 0x73A8C9, 0x60E27B, 0xC08C6B,
0N/A};
0N/A
0N/Astatic const int npio2_hw[] = {
0N/A 0x3FF921FB, 0x400921FB, 0x4012D97C, 0x401921FB, 0x401F6A7A, 0x4022D97C,
0N/A 0x4025FDBB, 0x402921FB, 0x402C463A, 0x402F6A7A, 0x4031475C, 0x4032D97C,
0N/A 0x40346B9C, 0x4035FDBB, 0x40378FDB, 0x403921FB, 0x403AB41B, 0x403C463A,
0N/A 0x403DD85A, 0x403F6A7A, 0x40407E4C, 0x4041475C, 0x4042106C, 0x4042D97C,
0N/A 0x4043A28C, 0x40446B9C, 0x404534AC, 0x4045FDBB, 0x4046C6CB, 0x40478FDB,
0N/A 0x404858EB, 0x404921FB,
0N/A};
0N/A
0N/A/*
0N/A * invpio2: 53 bits of 2/pi
0N/A * pio2_1: first 33 bit of pi/2
0N/A * pio2_1t: pi/2 - pio2_1
0N/A * pio2_2: second 33 bit of pi/2
0N/A * pio2_2t: pi/2 - (pio2_1+pio2_2)
0N/A * pio2_3: third 33 bit of pi/2
0N/A * pio2_3t: pi/2 - (pio2_1+pio2_2+pio2_3)
0N/A */
0N/A
0N/Astatic const double
0N/AzeroA = 0.00000000000000000000e+00, /* 0x00000000, 0x00000000 */
0N/Ahalf = 5.00000000000000000000e-01, /* 0x3FE00000, 0x00000000 */
0N/Atwo24A = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */
0N/Ainvpio2 = 6.36619772367581382433e-01, /* 0x3FE45F30, 0x6DC9C883 */
0N/Apio2_1 = 1.57079632673412561417e+00, /* 0x3FF921FB, 0x54400000 */
0N/Apio2_1t = 6.07710050650619224932e-11, /* 0x3DD0B461, 0x1A626331 */
0N/Apio2_2 = 6.07710050630396597660e-11, /* 0x3DD0B461, 0x1A600000 */
0N/Apio2_2t = 2.02226624879595063154e-21, /* 0x3BA3198A, 0x2E037073 */
0N/Apio2_3 = 2.02226624871116645580e-21, /* 0x3BA3198A, 0x2E000000 */
0N/Apio2_3t = 8.47842766036889956997e-32; /* 0x397B839A, 0x252049C1 */
0N/A
1405N/Astatic SAFEBUF int __ieee754_rem_pio2(double x, double *y) {
0N/A double z,w,t,r,fn;
0N/A double tx[3];
0N/A int e0,i,j,nx,n,ix,hx,i0;
0N/A
0N/A i0 = ((*(int*)&two24A)>>30)^1; /* high word index */
0N/A hx = *(i0+(int*)&x); /* high word of x */
0N/A ix = hx&0x7fffffff;
0N/A if(ix<=0x3fe921fb) /* |x| ~<= pi/4 , no need for reduction */
0N/A {y[0] = x; y[1] = 0; return 0;}
0N/A if(ix<0x4002d97c) { /* |x| < 3pi/4, special case with n=+-1 */
0N/A if(hx>0) {
0N/A z = x - pio2_1;
0N/A if(ix!=0x3ff921fb) { /* 33+53 bit pi is good enough */
0N/A y[0] = z - pio2_1t;
0N/A y[1] = (z-y[0])-pio2_1t;
0N/A } else { /* near pi/2, use 33+33+53 bit pi */
0N/A z -= pio2_2;
0N/A y[0] = z - pio2_2t;
0N/A y[1] = (z-y[0])-pio2_2t;
0N/A }
0N/A return 1;
0N/A } else { /* negative x */
0N/A z = x + pio2_1;
0N/A if(ix!=0x3ff921fb) { /* 33+53 bit pi is good enough */
0N/A y[0] = z + pio2_1t;
0N/A y[1] = (z-y[0])+pio2_1t;
0N/A } else { /* near pi/2, use 33+33+53 bit pi */
0N/A z += pio2_2;
0N/A y[0] = z + pio2_2t;
0N/A y[1] = (z-y[0])+pio2_2t;
0N/A }
0N/A return -1;
0N/A }
0N/A }
0N/A if(ix<=0x413921fb) { /* |x| ~<= 2^19*(pi/2), medium size */
0N/A t = fabsd(x);
0N/A n = (int) (t*invpio2+half);
0N/A fn = (double)n;
0N/A r = t-fn*pio2_1;
0N/A w = fn*pio2_1t; /* 1st round good to 85 bit */
0N/A if(n<32&&ix!=npio2_hw[n-1]) {
0N/A y[0] = r-w; /* quick check no cancellation */
0N/A } else {
0N/A j = ix>>20;
0N/A y[0] = r-w;
0N/A i = j-(((*(i0+(int*)&y[0]))>>20)&0x7ff);
0N/A if(i>16) { /* 2nd iteration needed, good to 118 */
0N/A t = r;
0N/A w = fn*pio2_2;
0N/A r = t-w;
0N/A w = fn*pio2_2t-((t-r)-w);
0N/A y[0] = r-w;
0N/A i = j-(((*(i0+(int*)&y[0]))>>20)&0x7ff);
0N/A if(i>49) { /* 3rd iteration need, 151 bits acc */
0N/A t = r; /* will cover all possible cases */
0N/A w = fn*pio2_3;
0N/A r = t-w;
0N/A w = fn*pio2_3t-((t-r)-w);
0N/A y[0] = r-w;
0N/A }
0N/A }
0N/A }
0N/A y[1] = (r-y[0])-w;
0N/A if(hx<0) {y[0] = -y[0]; y[1] = -y[1]; return -n;}
0N/A else return n;
0N/A }
0N/A /*
0N/A * all other (large) arguments
0N/A */
0N/A if(ix>=0x7ff00000) { /* x is inf or NaN */
0N/A y[0]=y[1]=x-x; return 0;
0N/A }
0N/A /* set z = scalbn(|x|,ilogb(x)-23) */
0N/A *(1-i0+(int*)&z) = *(1-i0+(int*)&x);
0N/A e0 = (ix>>20)-1046; /* e0 = ilogb(z)-23; */
0N/A *(i0+(int*)&z) = ix - (e0<<20);
0N/A for(i=0;i<2;i++) {
0N/A tx[i] = (double)((int)(z));
0N/A z = (z-tx[i])*two24A;
0N/A }
0N/A tx[2] = z;
0N/A nx = 3;
0N/A while(tx[nx-1]==zeroA) nx--; /* skip zero term */
0N/A n = __kernel_rem_pio2(tx,y,e0,nx,2,two_over_pi);
0N/A if(hx<0) {y[0] = -y[0]; y[1] = -y[1]; return -n;}
0N/A return n;
0N/A}
0N/A
0N/A
0N/A/* __kernel_sin( x, y, iy)
0N/A * kernel sin function on [-pi/4, pi/4], pi/4 ~ 0.7854
0N/A * Input x is assumed to be bounded by ~pi/4 in magnitude.
0N/A * Input y is the tail of x.
0N/A * Input iy indicates whether y is 0. (if iy=0, y assume to be 0).
0N/A *
0N/A * Algorithm
0N/A * 1. Since sin(-x) = -sin(x), we need only to consider positive x.
0N/A * 2. if x < 2^-27 (hx<0x3e400000 0), return x with inexact if x!=0.
0N/A * 3. sin(x) is approximated by a polynomial of degree 13 on
0N/A * [0,pi/4]
0N/A * 3 13
0N/A * sin(x) ~ x + S1*x + ... + S6*x
0N/A * where
0N/A *
0N/A * |sin(x) 2 4 6 8 10 12 | -58
0N/A * |----- - (1+S1*x +S2*x +S3*x +S4*x +S5*x +S6*x )| <= 2
0N/A * | x |
0N/A *
0N/A * 4. sin(x+y) = sin(x) + sin'(x')*y
0N/A * ~ sin(x) + (1-x*x/2)*y
0N/A * For better accuracy, let
0N/A * 3 2 2 2 2
0N/A * r = x *(S2+x *(S3+x *(S4+x *(S5+x *S6))))
0N/A * then 3 2
0N/A * sin(x) = x + (S1*x + (x *(r-y/2)+y))
0N/A */
0N/A
0N/Astatic const double
0N/AS1 = -1.66666666666666324348e-01, /* 0xBFC55555, 0x55555549 */
0N/AS2 = 8.33333333332248946124e-03, /* 0x3F811111, 0x1110F8A6 */
0N/AS3 = -1.98412698298579493134e-04, /* 0xBF2A01A0, 0x19C161D5 */
0N/AS4 = 2.75573137070700676789e-06, /* 0x3EC71DE3, 0x57B1FE7D */
0N/AS5 = -2.50507602534068634195e-08, /* 0xBE5AE5E6, 0x8A2B9CEB */
0N/AS6 = 1.58969099521155010221e-10; /* 0x3DE5D93A, 0x5ACFD57C */
0N/A
0N/Astatic double __kernel_sin(double x, double y, int iy)
0N/A{
0N/A double z,r,v;
0N/A int ix;
0N/A ix = __HI(x)&0x7fffffff; /* high word of x */
0N/A if(ix<0x3e400000) /* |x| < 2**-27 */
0N/A {if((int)x==0) return x;} /* generate inexact */
0N/A z = x*x;
0N/A v = z*x;
0N/A r = S2+z*(S3+z*(S4+z*(S5+z*S6)));
0N/A if(iy==0) return x+v*(S1+z*r);
0N/A else return x-((z*(half*y-v*r)-y)-v*S1);
0N/A}
0N/A
0N/A/*
0N/A * __kernel_cos( x, y )
0N/A * kernel cos function on [-pi/4, pi/4], pi/4 ~ 0.785398164
0N/A * Input x is assumed to be bounded by ~pi/4 in magnitude.
0N/A * Input y is the tail of x.
0N/A *
0N/A * Algorithm
0N/A * 1. Since cos(-x) = cos(x), we need only to consider positive x.
0N/A * 2. if x < 2^-27 (hx<0x3e400000 0), return 1 with inexact if x!=0.
0N/A * 3. cos(x) is approximated by a polynomial of degree 14 on
0N/A * [0,pi/4]
0N/A * 4 14
0N/A * cos(x) ~ 1 - x*x/2 + C1*x + ... + C6*x
0N/A * where the remez error is
0N/A *
0N/A * | 2 4 6 8 10 12 14 | -58
0N/A * |cos(x)-(1-.5*x +C1*x +C2*x +C3*x +C4*x +C5*x +C6*x )| <= 2
0N/A * | |
0N/A *
0N/A * 4 6 8 10 12 14
0N/A * 4. let r = C1*x +C2*x +C3*x +C4*x +C5*x +C6*x , then
0N/A * cos(x) = 1 - x*x/2 + r
0N/A * since cos(x+y) ~ cos(x) - sin(x)*y
0N/A * ~ cos(x) - x*y,
0N/A * a correction term is necessary in cos(x) and hence
0N/A * cos(x+y) = 1 - (x*x/2 - (r - x*y))
0N/A * For better accuracy when x > 0.3, let qx = |x|/4 with
0N/A * the last 32 bits mask off, and if x > 0.78125, let qx = 0.28125.
0N/A * Then
0N/A * cos(x+y) = (1-qx) - ((x*x/2-qx) - (r-x*y)).
0N/A * Note that 1-qx and (x*x/2-qx) is EXACT here, and the
0N/A * magnitude of the latter is at least a quarter of x*x/2,
0N/A * thus, reducing the rounding error in the subtraction.
0N/A */
0N/A
0N/Astatic const double
0N/AC1 = 4.16666666666666019037e-02, /* 0x3FA55555, 0x5555554C */
0N/AC2 = -1.38888888888741095749e-03, /* 0xBF56C16C, 0x16C15177 */
0N/AC3 = 2.48015872894767294178e-05, /* 0x3EFA01A0, 0x19CB1590 */
0N/AC4 = -2.75573143513906633035e-07, /* 0xBE927E4F, 0x809C52AD */
0N/AC5 = 2.08757232129817482790e-09, /* 0x3E21EE9E, 0xBDB4B1C4 */
0N/AC6 = -1.13596475577881948265e-11; /* 0xBDA8FAE9, 0xBE8838D4 */
0N/A
0N/Astatic double __kernel_cos(double x, double y)
0N/A{
0N/A double a,hz,z,r,qx;
0N/A int ix;
0N/A ix = __HI(x)&0x7fffffff; /* ix = |x|'s high word*/
0N/A if(ix<0x3e400000) { /* if x < 2**27 */
0N/A if(((int)x)==0) return one; /* generate inexact */
0N/A }
0N/A z = x*x;
0N/A r = z*(C1+z*(C2+z*(C3+z*(C4+z*(C5+z*C6)))));
0N/A if(ix < 0x3FD33333) /* if |x| < 0.3 */
0N/A return one - (0.5*z - (z*r - x*y));
0N/A else {
0N/A if(ix > 0x3fe90000) { /* x > 0.78125 */
0N/A qx = 0.28125;
0N/A } else {
0N/A __HI(qx) = ix-0x00200000; /* x/4 */
0N/A __LO(qx) = 0;
0N/A }
0N/A hz = 0.5*z-qx;
0N/A a = one-qx;
0N/A return a - (hz - (z*r-x*y));
0N/A }
0N/A}
0N/A
0N/A/* __kernel_tan( x, y, k )
0N/A * kernel tan function on [-pi/4, pi/4], pi/4 ~ 0.7854
0N/A * Input x is assumed to be bounded by ~pi/4 in magnitude.
0N/A * Input y is the tail of x.
0N/A * Input k indicates whether tan (if k=1) or
0N/A * -1/tan (if k= -1) is returned.
0N/A *
0N/A * Algorithm
0N/A * 1. Since tan(-x) = -tan(x), we need only to consider positive x.
0N/A * 2. if x < 2^-28 (hx<0x3e300000 0), return x with inexact if x!=0.
0N/A * 3. tan(x) is approximated by a odd polynomial of degree 27 on
0N/A * [0,0.67434]
0N/A * 3 27
0N/A * tan(x) ~ x + T1*x + ... + T13*x
0N/A * where
0N/A *
0N/A * |tan(x) 2 4 26 | -59.2
0N/A * |----- - (1+T1*x +T2*x +.... +T13*x )| <= 2
0N/A * | x |
0N/A *
0N/A * Note: tan(x+y) = tan(x) + tan'(x)*y
0N/A * ~ tan(x) + (1+x*x)*y
0N/A * Therefore, for better accuracy in computing tan(x+y), let
0N/A * 3 2 2 2 2
0N/A * r = x *(T2+x *(T3+x *(...+x *(T12+x *T13))))
0N/A * then
0N/A * 3 2
0N/A * tan(x+y) = x + (T1*x + (x *(r+y)+y))
0N/A *
0N/A * 4. For x in [0.67434,pi/4], let y = pi/4 - x, then
0N/A * tan(x) = tan(pi/4-y) = (1-tan(y))/(1+tan(y))
0N/A * = 1 - 2*(tan(y) - (tan(y)^2)/(1+tan(y)))
0N/A */
0N/A
0N/Astatic const double
0N/Apio4 = 7.85398163397448278999e-01, /* 0x3FE921FB, 0x54442D18 */
0N/Apio4lo= 3.06161699786838301793e-17, /* 0x3C81A626, 0x33145C07 */
0N/AT[] = {
0N/A 3.33333333333334091986e-01, /* 0x3FD55555, 0x55555563 */
0N/A 1.33333333333201242699e-01, /* 0x3FC11111, 0x1110FE7A */
0N/A 5.39682539762260521377e-02, /* 0x3FABA1BA, 0x1BB341FE */
0N/A 2.18694882948595424599e-02, /* 0x3F9664F4, 0x8406D637 */
0N/A 8.86323982359930005737e-03, /* 0x3F8226E3, 0xE96E8493 */
0N/A 3.59207910759131235356e-03, /* 0x3F6D6D22, 0xC9560328 */
0N/A 1.45620945432529025516e-03, /* 0x3F57DBC8, 0xFEE08315 */
0N/A 5.88041240820264096874e-04, /* 0x3F4344D8, 0xF2F26501 */
0N/A 2.46463134818469906812e-04, /* 0x3F3026F7, 0x1A8D1068 */
0N/A 7.81794442939557092300e-05, /* 0x3F147E88, 0xA03792A6 */
0N/A 7.14072491382608190305e-05, /* 0x3F12B80F, 0x32F0A7E9 */
0N/A -1.85586374855275456654e-05, /* 0xBEF375CB, 0xDB605373 */
0N/A 2.59073051863633712884e-05, /* 0x3EFB2A70, 0x74BF7AD4 */
0N/A};
0N/A
0N/Astatic double __kernel_tan(double x, double y, int iy)
0N/A{
0N/A double z,r,v,w,s;
0N/A int ix,hx;
0N/A hx = __HI(x); /* high word of x */
0N/A ix = hx&0x7fffffff; /* high word of |x| */
0N/A if(ix<0x3e300000) { /* x < 2**-28 */
0N/A if((int)x==0) { /* generate inexact */
0N/A if (((ix | __LO(x)) | (iy + 1)) == 0)
0N/A return one / fabsd(x);
0N/A else {
0N/A if (iy == 1)
0N/A return x;
0N/A else { /* compute -1 / (x+y) carefully */
0N/A double a, t;
0N/A
0N/A z = w = x + y;
0N/A __LO(z) = 0;
0N/A v = y - (z - x);
0N/A t = a = -one / w;
0N/A __LO(t) = 0;
0N/A s = one + t * z;
0N/A return t + a * (s + t * v);
0N/A }
0N/A }
0N/A }
0N/A }
0N/A if(ix>=0x3FE59428) { /* |x|>=0.6744 */
0N/A if(hx<0) {x = -x; y = -y;}
0N/A z = pio4-x;
0N/A w = pio4lo-y;
0N/A x = z+w; y = 0.0;
0N/A }
0N/A z = x*x;
0N/A w = z*z;
0N/A /* Break x^5*(T[1]+x^2*T[2]+...) into
0N/A * x^5(T[1]+x^4*T[3]+...+x^20*T[11]) +
0N/A * x^5(x^2*(T[2]+x^4*T[4]+...+x^22*[T12]))
0N/A */
0N/A r = T[1]+w*(T[3]+w*(T[5]+w*(T[7]+w*(T[9]+w*T[11]))));
0N/A v = z*(T[2]+w*(T[4]+w*(T[6]+w*(T[8]+w*(T[10]+w*T[12])))));
0N/A s = z*x;
0N/A r = y + z*(s*(r+v)+y);
0N/A r += T[0]*s;
0N/A w = x+r;
0N/A if(ix>=0x3FE59428) {
0N/A v = (double)iy;
0N/A return (double)(1-((hx>>30)&2))*(v-2.0*(x-(w*w/(w+v)-r)));
0N/A }
0N/A if(iy==1) return w;
0N/A else { /* if allow error up to 2 ulp,
0N/A simply return -1.0/(x+r) here */
0N/A /* compute -1.0/(x+r) accurately */
0N/A double a,t;
0N/A z = w;
0N/A __LO(z) = 0;
0N/A v = r-(z - x); /* z+v = r+x */
0N/A t = a = -1.0/w; /* a = -1.0/w */
0N/A __LO(t) = 0;
0N/A s = 1.0+t*z;
0N/A return t+a*(s+t*v);
0N/A }
0N/A}
0N/A
0N/A
0N/A//----------------------------------------------------------------------
0N/A//
0N/A// Routines for new sin/cos implementation
0N/A//
0N/A//----------------------------------------------------------------------
0N/A
0N/A/* sin(x)
0N/A * Return sine function of x.
0N/A *
0N/A * kernel function:
0N/A * __kernel_sin ... sine function on [-pi/4,pi/4]
0N/A * __kernel_cos ... cose function on [-pi/4,pi/4]
0N/A * __ieee754_rem_pio2 ... argument reduction routine
0N/A *
0N/A * Method.
0N/A * Let S,C and T denote the sin, cos and tan respectively on
0N/A * [-PI/4, +PI/4]. Reduce the argument x to y1+y2 = x-k*pi/2
0N/A * in [-pi/4 , +pi/4], and let n = k mod 4.
0N/A * We have
0N/A *
0N/A * n sin(x) cos(x) tan(x)
0N/A * ----------------------------------------------------------
0N/A * 0 S C T
0N/A * 1 C -S -1/T
0N/A * 2 -S -C T
0N/A * 3 -C S -1/T
0N/A * ----------------------------------------------------------
0N/A *
0N/A * Special cases:
0N/A * Let trig be any of sin, cos, or tan.
0N/A * trig(+-INF) is NaN, with signals;
0N/A * trig(NaN) is that NaN;
0N/A *
0N/A * Accuracy:
0N/A * TRIG(x) returns trig(x) nearly rounded
0N/A */
0N/A
0N/AJRT_LEAF(jdouble, SharedRuntime::dsin(jdouble x))
0N/A double y[2],z=0.0;
0N/A int n, ix;
0N/A
0N/A /* High word of x. */
0N/A ix = __HI(x);
0N/A
0N/A /* |x| ~< pi/4 */
0N/A ix &= 0x7fffffff;
0N/A if(ix <= 0x3fe921fb) return __kernel_sin(x,z,0);
0N/A
0N/A /* sin(Inf or NaN) is NaN */
0N/A else if (ix>=0x7ff00000) return x-x;
0N/A
0N/A /* argument reduction needed */
0N/A else {
0N/A n = __ieee754_rem_pio2(x,y);
0N/A switch(n&3) {
0N/A case 0: return __kernel_sin(y[0],y[1],1);
0N/A case 1: return __kernel_cos(y[0],y[1]);
0N/A case 2: return -__kernel_sin(y[0],y[1],1);
0N/A default:
0N/A return -__kernel_cos(y[0],y[1]);
0N/A }
0N/A }
0N/AJRT_END
0N/A
0N/A/* cos(x)
0N/A * Return cosine function of x.
0N/A *
0N/A * kernel function:
0N/A * __kernel_sin ... sine function on [-pi/4,pi/4]
0N/A * __kernel_cos ... cosine function on [-pi/4,pi/4]
0N/A * __ieee754_rem_pio2 ... argument reduction routine
0N/A *
0N/A * Method.
0N/A * Let S,C and T denote the sin, cos and tan respectively on
0N/A * [-PI/4, +PI/4]. Reduce the argument x to y1+y2 = x-k*pi/2
0N/A * in [-pi/4 , +pi/4], and let n = k mod 4.
0N/A * We have
0N/A *
0N/A * n sin(x) cos(x) tan(x)
0N/A * ----------------------------------------------------------
0N/A * 0 S C T
0N/A * 1 C -S -1/T
0N/A * 2 -S -C T
0N/A * 3 -C S -1/T
0N/A * ----------------------------------------------------------
0N/A *
0N/A * Special cases:
0N/A * Let trig be any of sin, cos, or tan.
0N/A * trig(+-INF) is NaN, with signals;
0N/A * trig(NaN) is that NaN;
0N/A *
0N/A * Accuracy:
0N/A * TRIG(x) returns trig(x) nearly rounded
0N/A */
0N/A
0N/AJRT_LEAF(jdouble, SharedRuntime::dcos(jdouble x))
0N/A double y[2],z=0.0;
0N/A int n, ix;
0N/A
0N/A /* High word of x. */
0N/A ix = __HI(x);
0N/A
0N/A /* |x| ~< pi/4 */
0N/A ix &= 0x7fffffff;
0N/A if(ix <= 0x3fe921fb) return __kernel_cos(x,z);
0N/A
0N/A /* cos(Inf or NaN) is NaN */
0N/A else if (ix>=0x7ff00000) return x-x;
0N/A
0N/A /* argument reduction needed */
0N/A else {
0N/A n = __ieee754_rem_pio2(x,y);
0N/A switch(n&3) {
0N/A case 0: return __kernel_cos(y[0],y[1]);
0N/A case 1: return -__kernel_sin(y[0],y[1],1);
0N/A case 2: return -__kernel_cos(y[0],y[1]);
0N/A default:
0N/A return __kernel_sin(y[0],y[1],1);
0N/A }
0N/A }
0N/AJRT_END
0N/A
0N/A/* tan(x)
0N/A * Return tangent function of x.
0N/A *
0N/A * kernel function:
0N/A * __kernel_tan ... tangent function on [-pi/4,pi/4]
0N/A * __ieee754_rem_pio2 ... argument reduction routine
0N/A *
0N/A * Method.
0N/A * Let S,C and T denote the sin, cos and tan respectively on
0N/A * [-PI/4, +PI/4]. Reduce the argument x to y1+y2 = x-k*pi/2
0N/A * in [-pi/4 , +pi/4], and let n = k mod 4.
0N/A * We have
0N/A *
0N/A * n sin(x) cos(x) tan(x)
0N/A * ----------------------------------------------------------
0N/A * 0 S C T
0N/A * 1 C -S -1/T
0N/A * 2 -S -C T
0N/A * 3 -C S -1/T
0N/A * ----------------------------------------------------------
0N/A *
0N/A * Special cases:
0N/A * Let trig be any of sin, cos, or tan.
0N/A * trig(+-INF) is NaN, with signals;
0N/A * trig(NaN) is that NaN;
0N/A *
0N/A * Accuracy:
0N/A * TRIG(x) returns trig(x) nearly rounded
0N/A */
0N/A
0N/AJRT_LEAF(jdouble, SharedRuntime::dtan(jdouble x))
0N/A double y[2],z=0.0;
0N/A int n, ix;
0N/A
0N/A /* High word of x. */
0N/A ix = __HI(x);
0N/A
0N/A /* |x| ~< pi/4 */
0N/A ix &= 0x7fffffff;
0N/A if(ix <= 0x3fe921fb) return __kernel_tan(x,z,1);
0N/A
0N/A /* tan(Inf or NaN) is NaN */
0N/A else if (ix>=0x7ff00000) return x-x; /* NaN */
0N/A
0N/A /* argument reduction needed */
0N/A else {
0N/A n = __ieee754_rem_pio2(x,y);
0N/A return __kernel_tan(y[0],y[1],1-((n&1)<<1)); /* 1 -- n even
0N/A -1 -- n odd */
0N/A }
0N/AJRT_END
0N/A
0N/A
0N/A#ifdef WIN32
0N/A# pragma optimize ( "", on )
0N/A#endif