1176491Smarcel/* $NetBSD: fpu_add.c,v 1.4 2005/12/11 12:18:42 christos Exp $ */ 2176491Smarcel 3176491Smarcel/* 4176491Smarcel * Copyright (c) 1992, 1993 5176491Smarcel * The Regents of the University of California. All rights reserved. 6176491Smarcel * 7176491Smarcel * This software was developed by the Computer Systems Engineering group 8176491Smarcel * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and 9176491Smarcel * contributed to Berkeley. 10176491Smarcel * 11176491Smarcel * All advertising materials mentioning features or use of this software 12176491Smarcel * must display the following acknowledgement: 13176491Smarcel * This product includes software developed by the University of 14176491Smarcel * California, Lawrence Berkeley Laboratory. 15176491Smarcel * 16176491Smarcel * Redistribution and use in source and binary forms, with or without 17176491Smarcel * modification, are permitted provided that the following conditions 18176491Smarcel * are met: 19176491Smarcel * 1. Redistributions of source code must retain the above copyright 20176491Smarcel * notice, this list of conditions and the following disclaimer. 21176491Smarcel * 2. Redistributions in binary form must reproduce the above copyright 22176491Smarcel * notice, this list of conditions and the following disclaimer in the 23176491Smarcel * documentation and/or other materials provided with the distribution. 24176491Smarcel * 3. Neither the name of the University nor the names of its contributors 25176491Smarcel * may be used to endorse or promote products derived from this software 26176491Smarcel * without specific prior written permission. 27176491Smarcel * 28176491Smarcel * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 29176491Smarcel * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 30176491Smarcel * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 31176491Smarcel * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 32176491Smarcel * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 33176491Smarcel * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 34176491Smarcel * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 35176491Smarcel * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 36176491Smarcel * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 37176491Smarcel * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 38176491Smarcel * SUCH DAMAGE. 39176491Smarcel * 40176491Smarcel * @(#)fpu_add.c 8.1 (Berkeley) 6/11/93 41176491Smarcel */ 42176491Smarcel 43176491Smarcel/* 44176491Smarcel * Perform an FPU add (return x + y). 45176491Smarcel * 46176491Smarcel * To subtract, negate y and call add. 47176491Smarcel */ 48176491Smarcel 49176491Smarcel#include <sys/cdefs.h> 50176491Smarcel__FBSDID("$FreeBSD$"); 51176491Smarcel 52178030Sgrehan#include <sys/types.h> 53176491Smarcel#include <sys/systm.h> 54176491Smarcel 55176491Smarcel#include <machine/fpu.h> 56176491Smarcel#include <machine/ieeefp.h> 57176491Smarcel#include <machine/reg.h> 58176491Smarcel 59176491Smarcel#include <powerpc/fpu/fpu_arith.h> 60176491Smarcel#include <powerpc/fpu/fpu_emu.h> 61176491Smarcel 62176491Smarcelstruct fpn * 63176491Smarcelfpu_add(struct fpemu *fe) 64176491Smarcel{ 65176491Smarcel struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2, *r; 66176491Smarcel u_int r0, r1, r2, r3; 67176491Smarcel int rd; 68176491Smarcel 69176491Smarcel /* 70176491Smarcel * Put the `heavier' operand on the right (see fpu_emu.h). 71176491Smarcel * Then we will have one of the following cases, taken in the 72176491Smarcel * following order: 73176491Smarcel * 74176491Smarcel * - y = NaN. Implied: if only one is a signalling NaN, y is. 75176491Smarcel * The result is y. 76176491Smarcel * - y = Inf. Implied: x != NaN (is 0, number, or Inf: the NaN 77176491Smarcel * case was taken care of earlier). 78176491Smarcel * If x = -y, the result is NaN. Otherwise the result 79176491Smarcel * is y (an Inf of whichever sign). 80176491Smarcel * - y is 0. Implied: x = 0. 81176491Smarcel * If x and y differ in sign (one positive, one negative), 82176491Smarcel * the result is +0 except when rounding to -Inf. If same: 83176491Smarcel * +0 + +0 = +0; -0 + -0 = -0. 84176491Smarcel * - x is 0. Implied: y != 0. 85176491Smarcel * Result is y. 86176491Smarcel * - other. Implied: both x and y are numbers. 87176491Smarcel * Do addition a la Hennessey & Patterson. 88176491Smarcel */ 89176491Smarcel DPRINTF(FPE_REG, ("fpu_add:\n")); 90176491Smarcel DUMPFPN(FPE_REG, x); 91176491Smarcel DUMPFPN(FPE_REG, y); 92176491Smarcel DPRINTF(FPE_REG, ("=>\n")); 93176491Smarcel ORDER(x, y); 94176491Smarcel if (ISNAN(y)) { 95176491Smarcel fe->fe_cx |= FPSCR_VXSNAN; 96176491Smarcel DUMPFPN(FPE_REG, y); 97176491Smarcel return (y); 98176491Smarcel } 99176491Smarcel if (ISINF(y)) { 100176491Smarcel if (ISINF(x) && x->fp_sign != y->fp_sign) { 101176491Smarcel fe->fe_cx |= FPSCR_VXISI; 102176491Smarcel return (fpu_newnan(fe)); 103176491Smarcel } 104176491Smarcel DUMPFPN(FPE_REG, y); 105176491Smarcel return (y); 106176491Smarcel } 107176491Smarcel rd = ((fe->fe_fpscr) & FPSCR_RN); 108176491Smarcel if (ISZERO(y)) { 109176491Smarcel if (rd != FP_RM) /* only -0 + -0 gives -0 */ 110176491Smarcel y->fp_sign &= x->fp_sign; 111176491Smarcel else /* any -0 operand gives -0 */ 112176491Smarcel y->fp_sign |= x->fp_sign; 113176491Smarcel DUMPFPN(FPE_REG, y); 114176491Smarcel return (y); 115176491Smarcel } 116176491Smarcel if (ISZERO(x)) { 117176491Smarcel DUMPFPN(FPE_REG, y); 118176491Smarcel return (y); 119176491Smarcel } 120176491Smarcel /* 121176491Smarcel * We really have two numbers to add, although their signs may 122176491Smarcel * differ. Make the exponents match, by shifting the smaller 123176491Smarcel * number right (e.g., 1.011 => 0.1011) and increasing its 124176491Smarcel * exponent (2^3 => 2^4). Note that we do not alter the exponents 125176491Smarcel * of x and y here. 126176491Smarcel */ 127176491Smarcel r = &fe->fe_f3; 128176491Smarcel r->fp_class = FPC_NUM; 129176491Smarcel if (x->fp_exp == y->fp_exp) { 130176491Smarcel r->fp_exp = x->fp_exp; 131176491Smarcel r->fp_sticky = 0; 132176491Smarcel } else { 133176491Smarcel if (x->fp_exp < y->fp_exp) { 134176491Smarcel /* 135176491Smarcel * Try to avoid subtract case iii (see below). 136176491Smarcel * This also guarantees that x->fp_sticky = 0. 137176491Smarcel */ 138176491Smarcel SWAP(x, y); 139176491Smarcel } 140176491Smarcel /* now x->fp_exp > y->fp_exp */ 141176491Smarcel r->fp_exp = x->fp_exp; 142176491Smarcel r->fp_sticky = fpu_shr(y, x->fp_exp - y->fp_exp); 143176491Smarcel } 144176491Smarcel r->fp_sign = x->fp_sign; 145176491Smarcel if (x->fp_sign == y->fp_sign) { 146176491Smarcel FPU_DECL_CARRY 147176491Smarcel 148176491Smarcel /* 149176491Smarcel * The signs match, so we simply add the numbers. The result 150176491Smarcel * may be `supernormal' (as big as 1.111...1 + 1.111...1, or 151176491Smarcel * 11.111...0). If so, a single bit shift-right will fix it 152176491Smarcel * (but remember to adjust the exponent). 153176491Smarcel */ 154176491Smarcel /* r->fp_mant = x->fp_mant + y->fp_mant */ 155176491Smarcel FPU_ADDS(r->fp_mant[3], x->fp_mant[3], y->fp_mant[3]); 156176491Smarcel FPU_ADDCS(r->fp_mant[2], x->fp_mant[2], y->fp_mant[2]); 157176491Smarcel FPU_ADDCS(r->fp_mant[1], x->fp_mant[1], y->fp_mant[1]); 158176491Smarcel FPU_ADDC(r0, x->fp_mant[0], y->fp_mant[0]); 159176491Smarcel if ((r->fp_mant[0] = r0) >= FP_2) { 160176491Smarcel (void) fpu_shr(r, 1); 161176491Smarcel r->fp_exp++; 162176491Smarcel } 163176491Smarcel } else { 164176491Smarcel FPU_DECL_CARRY 165176491Smarcel 166176491Smarcel /* 167176491Smarcel * The signs differ, so things are rather more difficult. 168176491Smarcel * H&P would have us negate the negative operand and add; 169176491Smarcel * this is the same as subtracting the negative operand. 170176491Smarcel * This is quite a headache. Instead, we will subtract 171176491Smarcel * y from x, regardless of whether y itself is the negative 172176491Smarcel * operand. When this is done one of three conditions will 173176491Smarcel * hold, depending on the magnitudes of x and y: 174176491Smarcel * case i) |x| > |y|. The result is just x - y, 175176491Smarcel * with x's sign, but it may need to be normalized. 176176491Smarcel * case ii) |x| = |y|. The result is 0 (maybe -0) 177176491Smarcel * so must be fixed up. 178176491Smarcel * case iii) |x| < |y|. We goofed; the result should 179176491Smarcel * be (y - x), with the same sign as y. 180176491Smarcel * We could compare |x| and |y| here and avoid case iii, 181176491Smarcel * but that would take just as much work as the subtract. 182176491Smarcel * We can tell case iii has occurred by an overflow. 183176491Smarcel * 184176491Smarcel * N.B.: since x->fp_exp >= y->fp_exp, x->fp_sticky = 0. 185176491Smarcel */ 186176491Smarcel /* r->fp_mant = x->fp_mant - y->fp_mant */ 187176491Smarcel FPU_SET_CARRY(y->fp_sticky); 188176491Smarcel FPU_SUBCS(r3, x->fp_mant[3], y->fp_mant[3]); 189176491Smarcel FPU_SUBCS(r2, x->fp_mant[2], y->fp_mant[2]); 190176491Smarcel FPU_SUBCS(r1, x->fp_mant[1], y->fp_mant[1]); 191176491Smarcel FPU_SUBC(r0, x->fp_mant[0], y->fp_mant[0]); 192176491Smarcel if (r0 < FP_2) { 193176491Smarcel /* cases i and ii */ 194176491Smarcel if ((r0 | r1 | r2 | r3) == 0) { 195176491Smarcel /* case ii */ 196176491Smarcel r->fp_class = FPC_ZERO; 197176491Smarcel r->fp_sign = rd == FP_RM; 198176491Smarcel return (r); 199176491Smarcel } 200176491Smarcel } else { 201176491Smarcel /* 202176491Smarcel * Oops, case iii. This can only occur when the 203176491Smarcel * exponents were equal, in which case neither 204176491Smarcel * x nor y have sticky bits set. Flip the sign 205176491Smarcel * (to y's sign) and negate the result to get y - x. 206176491Smarcel */ 207176491Smarcel#ifdef DIAGNOSTIC 208176491Smarcel if (x->fp_exp != y->fp_exp || r->fp_sticky) 209176491Smarcel panic("fpu_add"); 210176491Smarcel#endif 211176491Smarcel r->fp_sign = y->fp_sign; 212176491Smarcel FPU_SUBS(r3, 0, r3); 213176491Smarcel FPU_SUBCS(r2, 0, r2); 214176491Smarcel FPU_SUBCS(r1, 0, r1); 215176491Smarcel FPU_SUBC(r0, 0, r0); 216176491Smarcel } 217176491Smarcel r->fp_mant[3] = r3; 218176491Smarcel r->fp_mant[2] = r2; 219176491Smarcel r->fp_mant[1] = r1; 220176491Smarcel r->fp_mant[0] = r0; 221176491Smarcel if (r0 < FP_1) 222176491Smarcel fpu_norm(r); 223176491Smarcel } 224176491Smarcel DUMPFPN(FPE_REG, r); 225176491Smarcel return (r); 226176491Smarcel} 227