Home | History | Annotate | Line # | Download | only in fpu
fpu_add.c revision 1.2
      1 /*	$NetBSD: fpu_add.c,v 1.2 1994/11/20 20:52:34 deraadt Exp $ */
      2 
      3 /*
      4  * Copyright (c) 1992, 1993
      5  *	The Regents of the University of California.  All rights reserved.
      6  *
      7  * This software was developed by the Computer Systems Engineering group
      8  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
      9  * contributed to Berkeley.
     10  *
     11  * All advertising materials mentioning features or use of this software
     12  * must display the following acknowledgement:
     13  *	This product includes software developed by the University of
     14  *	California, Lawrence Berkeley Laboratory.
     15  *
     16  * Redistribution and use in source and binary forms, with or without
     17  * modification, are permitted provided that the following conditions
     18  * are met:
     19  * 1. Redistributions of source code must retain the above copyright
     20  *    notice, this list of conditions and the following disclaimer.
     21  * 2. Redistributions in binary form must reproduce the above copyright
     22  *    notice, this list of conditions and the following disclaimer in the
     23  *    documentation and/or other materials provided with the distribution.
     24  * 3. All advertising materials mentioning features or use of this software
     25  *    must display the following acknowledgement:
     26  *	This product includes software developed by the University of
     27  *	California, Berkeley and its contributors.
     28  * 4. Neither the name of the University nor the names of its contributors
     29  *    may be used to endorse or promote products derived from this software
     30  *    without specific prior written permission.
     31  *
     32  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     33  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     34  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     35  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     36  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     37  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     38  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     39  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     40  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     41  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     42  * SUCH DAMAGE.
     43  *
     44  *	@(#)fpu_add.c	8.1 (Berkeley) 6/11/93
     45  */
     46 
     47 /*
     48  * Perform an FPU add (return x + y).
     49  *
     50  * To subtract, negate y and call add.
     51  */
     52 
     53 #include <sys/types.h>
     54 
     55 #include <machine/reg.h>
     56 
     57 #include <sparc/fpu/fpu_arith.h>
     58 #include <sparc/fpu/fpu_emu.h>
     59 
     60 struct fpn *
     61 fpu_add(fe)
     62 	register struct fpemu *fe;
     63 {
     64 	register struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2, *r;
     65 	register u_int r0, r1, r2, r3;
     66 	register int rd;
     67 
     68 	/*
     69 	 * Put the `heavier' operand on the right (see fpu_emu.h).
     70 	 * Then we will have one of the following cases, taken in the
     71 	 * following order:
     72 	 *
     73 	 *  - y = NaN.  Implied: if only one is a signalling NaN, y is.
     74 	 *	The result is y.
     75 	 *  - y = Inf.  Implied: x != NaN (is 0, number, or Inf: the NaN
     76 	 *    case was taken care of earlier).
     77 	 *	If x = -y, the result is NaN.  Otherwise the result
     78 	 *	is y (an Inf of whichever sign).
     79 	 *  - y is 0.  Implied: x = 0.
     80 	 *	If x and y differ in sign (one positive, one negative),
     81 	 *	the result is +0 except when rounding to -Inf.  If same:
     82 	 *	+0 + +0 = +0; -0 + -0 = -0.
     83 	 *  - x is 0.  Implied: y != 0.
     84 	 *	Result is y.
     85 	 *  - other.  Implied: both x and y are numbers.
     86 	 *	Do addition a la Hennessey & Patterson.
     87 	 */
     88 	ORDER(x, y);
     89 	if (ISNAN(y))
     90 		return (y);
     91 	if (ISINF(y)) {
     92 		if (ISINF(x) && x->fp_sign != y->fp_sign)
     93 			return (fpu_newnan(fe));
     94 		return (y);
     95 	}
     96 	rd = ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK);
     97 	if (ISZERO(y)) {
     98 		if (rd != FSR_RD_RM)	/* only -0 + -0 gives -0 */
     99 			y->fp_sign &= x->fp_sign;
    100 		else			/* any -0 operand gives -0 */
    101 			y->fp_sign |= x->fp_sign;
    102 		return (y);
    103 	}
    104 	if (ISZERO(x))
    105 		return (y);
    106 	/*
    107 	 * We really have two numbers to add, although their signs may
    108 	 * differ.  Make the exponents match, by shifting the smaller
    109 	 * number right (e.g., 1.011 => 0.1011) and increasing its
    110 	 * exponent (2^3 => 2^4).  Note that we do not alter the exponents
    111 	 * of x and y here.
    112 	 */
    113 	r = &fe->fe_f3;
    114 	r->fp_class = FPC_NUM;
    115 	if (x->fp_exp == y->fp_exp) {
    116 		r->fp_exp = x->fp_exp;
    117 		r->fp_sticky = 0;
    118 	} else {
    119 		if (x->fp_exp < y->fp_exp) {
    120 			/*
    121 			 * Try to avoid subtract case iii (see below).
    122 			 * This also guarantees that x->fp_sticky = 0.
    123 			 */
    124 			SWAP(x, y);
    125 		}
    126 		/* now x->fp_exp > y->fp_exp */
    127 		r->fp_exp = x->fp_exp;
    128 		r->fp_sticky = fpu_shr(y, x->fp_exp - y->fp_exp);
    129 	}
    130 	r->fp_sign = x->fp_sign;
    131 	if (x->fp_sign == y->fp_sign) {
    132 		FPU_DECL_CARRY
    133 
    134 		/*
    135 		 * The signs match, so we simply add the numbers.  The result
    136 		 * may be `supernormal' (as big as 1.111...1 + 1.111...1, or
    137 		 * 11.111...0).  If so, a single bit shift-right will fix it
    138 		 * (but remember to adjust the exponent).
    139 		 */
    140 		/* r->fp_mant = x->fp_mant + y->fp_mant */
    141 		FPU_ADDS(r->fp_mant[3], x->fp_mant[3], y->fp_mant[3]);
    142 		FPU_ADDCS(r->fp_mant[2], x->fp_mant[2], y->fp_mant[2]);
    143 		FPU_ADDCS(r->fp_mant[1], x->fp_mant[1], y->fp_mant[1]);
    144 		FPU_ADDC(r0, x->fp_mant[0], y->fp_mant[0]);
    145 		if ((r->fp_mant[0] = r0) >= FP_2) {
    146 			(void) fpu_shr(r, 1);
    147 			r->fp_exp++;
    148 		}
    149 	} else {
    150 		FPU_DECL_CARRY
    151 
    152 		/*
    153 		 * The signs differ, so things are rather more difficult.
    154 		 * H&P would have us negate the negative operand and add;
    155 		 * this is the same as subtracting the negative operand.
    156 		 * This is quite a headache.  Instead, we will subtract
    157 		 * y from x, regardless of whether y itself is the negative
    158 		 * operand.  When this is done one of three conditions will
    159 		 * hold, depending on the magnitudes of x and y:
    160 		 *   case i)   |x| > |y|.  The result is just x - y,
    161 		 *	with x's sign, but it may need to be normalized.
    162 		 *   case ii)  |x| = |y|.  The result is 0 (maybe -0)
    163 		 *	so must be fixed up.
    164 		 *   case iii) |x| < |y|.  We goofed; the result should
    165 		 *	be (y - x), with the same sign as y.
    166 		 * We could compare |x| and |y| here and avoid case iii,
    167 		 * but that would take just as much work as the subtract.
    168 		 * We can tell case iii has occurred by an overflow.
    169 		 *
    170 		 * N.B.: since x->fp_exp >= y->fp_exp, x->fp_sticky = 0.
    171 		 */
    172 		/* r->fp_mant = x->fp_mant - y->fp_mant */
    173 		FPU_SET_CARRY(y->fp_sticky);
    174 		FPU_SUBCS(r3, x->fp_mant[3], y->fp_mant[3]);
    175 		FPU_SUBCS(r2, x->fp_mant[2], y->fp_mant[2]);
    176 		FPU_SUBCS(r1, x->fp_mant[1], y->fp_mant[1]);
    177 		FPU_SUBC(r0, x->fp_mant[0], y->fp_mant[0]);
    178 		if (r0 < FP_2) {
    179 			/* cases i and ii */
    180 			if ((r0 | r1 | r2 | r3) == 0) {
    181 				/* case ii */
    182 				r->fp_class = FPC_ZERO;
    183 				r->fp_sign = rd == FSR_RD_RM;
    184 				return (r);
    185 			}
    186 		} else {
    187 			/*
    188 			 * Oops, case iii.  This can only occur when the
    189 			 * exponents were equal, in which case neither
    190 			 * x nor y have sticky bits set.  Flip the sign
    191 			 * (to y's sign) and negate the result to get y - x.
    192 			 */
    193 #ifdef DIAGNOSTIC
    194 			if (x->fp_exp != y->fp_exp || r->fp_sticky)
    195 				panic("fpu_add");
    196 #endif
    197 			r->fp_sign = y->fp_sign;
    198 			FPU_SUBS(r3, 0, r3);
    199 			FPU_SUBCS(r2, 0, r2);
    200 			FPU_SUBCS(r1, 0, r1);
    201 			FPU_SUBC(r0, 0, r0);
    202 		}
    203 		r->fp_mant[3] = r3;
    204 		r->fp_mant[2] = r2;
    205 		r->fp_mant[1] = r1;
    206 		r->fp_mant[0] = r0;
    207 		if (r0 < FP_1)
    208 			fpu_norm(r);
    209 	}
    210 	return (r);
    211 }
    212