1118611Snjl/* 2118611Snjl * Copyright (c) 1992, 1993 3118611Snjl * The Regents of the University of California. All rights reserved. 4151937Sjkim * 5118611Snjl * This software was developed by the Computer Systems Engineering group 6118611Snjl * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and 7118611Snjl * contributed to Berkeley. 8118611Snjl * 9118611Snjl * All advertising materials mentioning features or use of this software 10118611Snjl * must display the following acknowledgement: 11118611Snjl * This product includes software developed by the University of 12202771Sjkim * California, Lawrence Berkeley Laboratory. 13118611Snjl * 14118611Snjl * Redistribution and use in source and binary forms, with or without 15118611Snjl * modification, are permitted provided that the following conditions 16118611Snjl * are met: 17118611Snjl * 1. Redistributions of source code must retain the above copyright 18118611Snjl * notice, this list of conditions and the following disclaimer. 19118611Snjl * 2. Redistributions in binary form must reproduce the above copyright 20118611Snjl * notice, this list of conditions and the following disclaimer in the 21118611Snjl * documentation and/or other materials provided with the distribution. 22118611Snjl * 4. Neither the name of the University nor the names of its contributors 23118611Snjl * may be used to endorse or promote products derived from this software 24118611Snjl * without specific prior written permission. 25118611Snjl * 26118611Snjl * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27118611Snjl * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28118611Snjl * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29118611Snjl * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30118611Snjl * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31118611Snjl * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32118611Snjl * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33118611Snjl * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34118611Snjl * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35118611Snjl * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36118611Snjl * SUCH DAMAGE. 37118611Snjl * 38118611Snjl * @(#)fpu_subr.c 8.1 (Berkeley) 6/11/93 39118611Snjl * $NetBSD: fpu_subr.c,v 1.3 1996/03/14 19:42:01 christos Exp $ 40118611Snjl */ 41118611Snjl 42118611Snjl#include <sys/cdefs.h> 43118611Snjl__FBSDID("$FreeBSD$"); 44118611Snjl 45118611Snjl/* 46118611Snjl * FPU subroutines. 47118611Snjl */ 48118611Snjl 49118611Snjl#include <sys/param.h> 50118611Snjl 51118611Snjl#include <machine/frame.h> 52118611Snjl#include <machine/fp.h> 53118611Snjl#include <machine/fsr.h> 54118611Snjl#include <machine/instr.h> 55118611Snjl 56118611Snjl#include "fpu_arith.h" 57118611Snjl#include "fpu_emu.h" 58118611Snjl#include "fpu_extern.h" 59118611Snjl#include "__sparc_utrap_private.h" 60118611Snjl 61118611Snjl/* 62118611Snjl * Shift the given number right rsh bits. Any bits that `fall off' will get 63118611Snjl * shoved into the sticky field; we return the resulting sticky. Note that 64118611Snjl * shifting NaNs is legal (this will never shift all bits out); a NaN's 65118611Snjl * sticky field is ignored anyway. 66118611Snjl */ 67118611Snjlint 68118611Snjl__fpu_shr(struct fpn *fp, int rsh) 69118611Snjl{ 70118611Snjl u_int m0, m1, m2, m3, s; 71118611Snjl int lsh; 72118611Snjl 73118611Snjl#ifdef DIAGNOSTIC 74118611Snjl if (rsh <= 0 || (fp->fp_class != FPC_NUM && !ISNAN(fp))) 75118611Snjl __utrap_panic("fpu_rightshift 1"); 76118611Snjl#endif 77118611Snjl 78118611Snjl m0 = fp->fp_mant[0]; 79118611Snjl m1 = fp->fp_mant[1]; 80118611Snjl m2 = fp->fp_mant[2]; 81118611Snjl m3 = fp->fp_mant[3]; 82118611Snjl 83118611Snjl /* If shifting all the bits out, take a shortcut. */ 84118611Snjl if (rsh >= FP_NMANT) { 85118611Snjl#ifdef DIAGNOSTIC 86118611Snjl if ((m0 | m1 | m2 | m3) == 0) 87118611Snjl __utrap_panic("fpu_rightshift 2"); 88118611Snjl#endif 89118611Snjl fp->fp_mant[0] = 0; 90118611Snjl fp->fp_mant[1] = 0; 91118611Snjl fp->fp_mant[2] = 0; 92118611Snjl fp->fp_mant[3] = 0; 93118611Snjl#ifdef notdef 94118611Snjl if ((m0 | m1 | m2 | m3) == 0) 95118611Snjl fp->fp_class = FPC_ZERO; 96118611Snjl else 97118611Snjl#endif 98118611Snjl fp->fp_sticky = 1; 99118611Snjl return (1); 100118611Snjl } 101118611Snjl 102118611Snjl /* Squish out full words. */ 103118611Snjl s = fp->fp_sticky; 104118611Snjl if (rsh >= 32 * 3) { 105118611Snjl s |= m3 | m2 | m1; 106118611Snjl m3 = m0, m2 = 0, m1 = 0, m0 = 0; 107118611Snjl } else if (rsh >= 32 * 2) { 108118611Snjl s |= m3 | m2; 109118611Snjl m3 = m1, m2 = m0, m1 = 0, m0 = 0; 110118611Snjl } else if (rsh >= 32) { 111118611Snjl s |= m3; 112118611Snjl m3 = m2, m2 = m1, m1 = m0, m0 = 0; 113118611Snjl } 114118611Snjl 115118611Snjl /* Handle any remaining partial word. */ 116118611Snjl if ((rsh &= 31) != 0) { 117118611Snjl lsh = 32 - rsh; 118118611Snjl s |= m3 << lsh; 119118611Snjl m3 = (m3 >> rsh) | (m2 << lsh); 120118611Snjl m2 = (m2 >> rsh) | (m1 << lsh); 121118611Snjl m1 = (m1 >> rsh) | (m0 << lsh); 122118611Snjl m0 >>= rsh; 123118611Snjl } 124118611Snjl fp->fp_mant[0] = m0; 125118611Snjl fp->fp_mant[1] = m1; 126118611Snjl fp->fp_mant[2] = m2; 127118611Snjl fp->fp_mant[3] = m3; 128118611Snjl fp->fp_sticky = s; 129118611Snjl return (s); 130118611Snjl} 131118611Snjl 132118611Snjl/* 133118611Snjl * Force a number to be normal, i.e., make its fraction have all zero 134118611Snjl * bits before FP_1, then FP_1, then all 1 bits. This is used for denorms 135118611Snjl * and (sometimes) for intermediate results. 136118611Snjl * 137118611Snjl * Internally, this may use a `supernormal' -- a number whose fp_mant 138118611Snjl * is greater than or equal to 2.0 -- so as a side effect you can hand it 139118611Snjl * a supernormal and it will fix it (provided fp->fp_mant[3] == 0). 140118611Snjl */ 141193529Sjkimvoid 142193529Sjkim__fpu_norm(struct fpn *fp) 143193529Sjkim{ 144193529Sjkim u_int m0, m1, m2, m3, top, sup, nrm; 145118611Snjl int lsh, rsh, exp; 146151937Sjkim 147118611Snjl exp = fp->fp_exp; 148151937Sjkim m0 = fp->fp_mant[0]; 149151937Sjkim m1 = fp->fp_mant[1]; 150151937Sjkim m2 = fp->fp_mant[2]; 151118611Snjl m3 = fp->fp_mant[3]; 152118611Snjl 153118611Snjl /* Handle severe subnormals with 32-bit moves. */ 154118611Snjl if (m0 == 0) { 155118611Snjl if (m1) 156118611Snjl m0 = m1, m1 = m2, m2 = m3, m3 = 0, exp -= 32; 157118611Snjl else if (m2) 158118611Snjl m0 = m2, m1 = m3, m2 = 0, m3 = 0, exp -= 2 * 32; 159151937Sjkim else if (m3) 160151937Sjkim m0 = m3, m1 = 0, m2 = 0, m3 = 0, exp -= 3 * 32; 161151937Sjkim else { 162118611Snjl fp->fp_class = FPC_ZERO; 163118611Snjl return; 164118611Snjl } 165118611Snjl } 166118611Snjl 167118611Snjl /* Now fix any supernormal or remaining subnormal. */ 168118611Snjl nrm = FP_1; 169118611Snjl sup = nrm << 1; 170118611Snjl if (m0 >= sup) { 171118611Snjl /* 172118611Snjl * We have a supernormal number. We need to shift it right. 173118611Snjl * We may assume m3==0. 174118611Snjl */ 175118611Snjl for (rsh = 1, top = m0 >> 1; top >= sup; rsh++) /* XXX slow */ 176118611Snjl top >>= 1; 177118611Snjl exp += rsh; 178118611Snjl lsh = 32 - rsh; 179118611Snjl m3 = m2 << lsh; 180118611Snjl m2 = (m2 >> rsh) | (m1 << lsh); 181118611Snjl m1 = (m1 >> rsh) | (m0 << lsh); 182118611Snjl m0 = top; 183118611Snjl } else if (m0 < nrm) { 184118611Snjl /* 185118611Snjl * We have a regular denorm (a subnormal number), and need 186118611Snjl * to shift it left. 187118611Snjl */ 188118611Snjl for (lsh = 1, top = m0 << 1; top < nrm; lsh++) /* XXX slow */ 189118611Snjl top <<= 1; 190118611Snjl exp -= lsh; 191193529Sjkim rsh = 32 - lsh; 192193529Sjkim m0 = top | (m1 >> rsh); 193193529Sjkim m1 = (m1 << lsh) | (m2 >> rsh); 194193529Sjkim m2 = (m2 << lsh) | (m3 >> rsh); 195193529Sjkim m3 <<= lsh; 196193529Sjkim } 197118611Snjl 198193529Sjkim fp->fp_exp = exp; 199193529Sjkim fp->fp_mant[0] = m0; 200193529Sjkim fp->fp_mant[1] = m1; 201193529Sjkim fp->fp_mant[2] = m2; 202151937Sjkim fp->fp_mant[3] = m3; 203151937Sjkim} 204151937Sjkim 205118611Snjl/* 206118611Snjl * Concoct a `fresh' Quiet NaN per Appendix N. 207118611Snjl * As a side effect, we set NV (invalid) for the current exceptions. 208118611Snjl */ 209118611Snjlstruct fpn * 210118611Snjl__fpu_newnan(struct fpemu *fe) 211118611Snjl{ 212118611Snjl struct fpn *fp; 213151937Sjkim 214151937Sjkim fe->fe_cx = FSR_NV; 215151937Sjkim fp = &fe->fe_f3; 216151937Sjkim fp->fp_class = FPC_QNAN; 217118611Snjl fp->fp_sign = 0; 218151937Sjkim fp->fp_mant[0] = FP_1 - 1; 219151937Sjkim fp->fp_mant[1] = fp->fp_mant[2] = fp->fp_mant[3] = ~0; 220118611Snjl return (fp); 221151937Sjkim} 222151937Sjkim