xref: /csrg-svn/sys/sparc/fpu/fpu_add.c (revision 63319)
155111Storek /*
2*63319Sbostic  * Copyright (c) 1992, 1993
3*63319Sbostic  *	The Regents of the University of California.  All rights reserved.
455111Storek  *
555111Storek  * This software was developed by the Computer Systems Engineering group
655111Storek  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
755111Storek  * contributed to Berkeley.
855111Storek  *
955500Sbostic  * All advertising materials mentioning features or use of this software
1055500Sbostic  * must display the following acknowledgement:
1155500Sbostic  *	This product includes software developed by the University of
1259195Storek  *	California, Lawrence Berkeley Laboratory.
1355500Sbostic  *
1455111Storek  * %sccs.include.redist.c%
1555111Storek  *
16*63319Sbostic  *	@(#)fpu_add.c	8.1 (Berkeley) 06/11/93
1755111Storek  *
1859195Storek  * from: $Header: fpu_add.c,v 1.4 92/11/26 01:39:46 torek Exp $
1955111Storek  */
2055111Storek 
2155111Storek /*
2255111Storek  * Perform an FPU add (return x + y).
2355111Storek  *
2455111Storek  * To subtract, negate y and call add.
2555111Storek  */
2655111Storek 
2756537Sbostic #include <sys/types.h>
2855111Storek 
2956537Sbostic #include <machine/reg.h>
3055111Storek 
3156537Sbostic #include <sparc/fpu/fpu_arith.h>
3256537Sbostic #include <sparc/fpu/fpu_emu.h>
3355111Storek 
3455111Storek struct fpn *
fpu_add(fe)3555111Storek fpu_add(fe)
3655111Storek 	register struct fpemu *fe;
3755111Storek {
3855111Storek 	register struct fpn *x = &fe->fe_f1, *y = &fe->fe_f2, *r;
3955111Storek 	register u_int r0, r1, r2, r3;
4055111Storek 	register int rd;
4155111Storek 
4255111Storek 	/*
4355111Storek 	 * Put the `heavier' operand on the right (see fpu_emu.h).
4455111Storek 	 * Then we will have one of the following cases, taken in the
4555111Storek 	 * following order:
4655111Storek 	 *
4755111Storek 	 *  - y = NaN.  Implied: if only one is a signalling NaN, y is.
4855111Storek 	 *	The result is y.
4955111Storek 	 *  - y = Inf.  Implied: x != NaN (is 0, number, or Inf: the NaN
5055111Storek 	 *    case was taken care of earlier).
5155111Storek 	 *	If x = -y, the result is NaN.  Otherwise the result
5255111Storek 	 *	is y (an Inf of whichever sign).
5355111Storek 	 *  - y is 0.  Implied: x = 0.
5455111Storek 	 *	If x and y differ in sign (one positive, one negative),
5555111Storek 	 *	the result is +0 except when rounding to -Inf.  If same:
5655111Storek 	 *	+0 + +0 = +0; -0 + -0 = -0.
5755111Storek 	 *  - x is 0.  Implied: y != 0.
5855111Storek 	 *	Result is y.
5955111Storek 	 *  - other.  Implied: both x and y are numbers.
6055111Storek 	 *	Do addition a la Hennessey & Patterson.
6155111Storek 	 */
6255111Storek 	ORDER(x, y);
6355111Storek 	if (ISNAN(y))
6455111Storek 		return (y);
6555111Storek 	if (ISINF(y)) {
6655111Storek 		if (ISINF(x) && x->fp_sign != y->fp_sign)
6755111Storek 			return (fpu_newnan(fe));
6855111Storek 		return (y);
6955111Storek 	}
7055111Storek 	rd = ((fe->fe_fsr >> FSR_RD_SHIFT) & FSR_RD_MASK);
7155111Storek 	if (ISZERO(y)) {
7255111Storek 		if (rd != FSR_RD_RM)	/* only -0 + -0 gives -0 */
7355111Storek 			y->fp_sign &= x->fp_sign;
7455111Storek 		else			/* any -0 operand gives -0 */
7555111Storek 			y->fp_sign |= x->fp_sign;
7655111Storek 		return (y);
7755111Storek 	}
7855111Storek 	if (ISZERO(x))
7955111Storek 		return (y);
8055111Storek 	/*
8155111Storek 	 * We really have two numbers to add, although their signs may
8255111Storek 	 * differ.  Make the exponents match, by shifting the smaller
8355111Storek 	 * number right (e.g., 1.011 => 0.1011) and increasing its
8455111Storek 	 * exponent (2^3 => 2^4).  Note that we do not alter the exponents
8555111Storek 	 * of x and y here.
8655111Storek 	 */
8755111Storek 	r = &fe->fe_f3;
8855111Storek 	r->fp_class = FPC_NUM;
8955111Storek 	if (x->fp_exp == y->fp_exp) {
9055111Storek 		r->fp_exp = x->fp_exp;
9155111Storek 		r->fp_sticky = 0;
9255111Storek 	} else {
9355111Storek 		if (x->fp_exp < y->fp_exp) {
9455111Storek 			/*
9555111Storek 			 * Try to avoid subtract case iii (see below).
9655111Storek 			 * This also guarantees that x->fp_sticky = 0.
9755111Storek 			 */
9855111Storek 			SWAP(x, y);
9955111Storek 		}
10055111Storek 		/* now x->fp_exp > y->fp_exp */
10155111Storek 		r->fp_exp = x->fp_exp;
10255111Storek 		r->fp_sticky = fpu_shr(y, x->fp_exp - y->fp_exp);
10355111Storek 	}
10455111Storek 	r->fp_sign = x->fp_sign;
10555111Storek 	if (x->fp_sign == y->fp_sign) {
10655111Storek 		FPU_DECL_CARRY
10755111Storek 
10855111Storek 		/*
10955111Storek 		 * The signs match, so we simply add the numbers.  The result
11055111Storek 		 * may be `supernormal' (as big as 1.111...1 + 1.111...1, or
11155111Storek 		 * 11.111...0).  If so, a single bit shift-right will fix it
11255111Storek 		 * (but remember to adjust the exponent).
11355111Storek 		 */
11455111Storek 		/* r->fp_mant = x->fp_mant + y->fp_mant */
11555111Storek 		FPU_ADDS(r->fp_mant[3], x->fp_mant[3], y->fp_mant[3]);
11655111Storek 		FPU_ADDCS(r->fp_mant[2], x->fp_mant[2], y->fp_mant[2]);
11755111Storek 		FPU_ADDCS(r->fp_mant[1], x->fp_mant[1], y->fp_mant[1]);
11855111Storek 		FPU_ADDC(r0, x->fp_mant[0], y->fp_mant[0]);
11955111Storek 		if ((r->fp_mant[0] = r0) >= FP_2) {
12055111Storek 			(void) fpu_shr(r, 1);
12155111Storek 			r->fp_exp++;
12255111Storek 		}
12355111Storek 	} else {
12455111Storek 		FPU_DECL_CARRY
12555111Storek 
12655111Storek 		/*
12755111Storek 		 * The signs differ, so things are rather more difficult.
12855111Storek 		 * H&P would have us negate the negative operand and add;
12955111Storek 		 * this is the same as subtracting the negative operand.
13055111Storek 		 * This is quite a headache.  Instead, we will subtract
13155111Storek 		 * y from x, regardless of whether y itself is the negative
13255111Storek 		 * operand.  When this is done one of three conditions will
13355111Storek 		 * hold, depending on the magnitudes of x and y:
13455111Storek 		 *   case i)   |x| > |y|.  The result is just x - y,
13555111Storek 		 *	with x's sign, but it may need to be normalized.
13655111Storek 		 *   case ii)  |x| = |y|.  The result is 0 (maybe -0)
13755111Storek 		 *	so must be fixed up.
13855111Storek 		 *   case iii) |x| < |y|.  We goofed; the result should
13955111Storek 		 *	be (y - x), with the same sign as y.
14055111Storek 		 * We could compare |x| and |y| here and avoid case iii,
14155111Storek 		 * but that would take just as much work as the subtract.
14255111Storek 		 * We can tell case iii has occurred by an overflow.
14355111Storek 		 *
14455111Storek 		 * N.B.: since x->fp_exp >= y->fp_exp, x->fp_sticky = 0.
14555111Storek 		 */
14655111Storek 		/* r->fp_mant = x->fp_mant - y->fp_mant */
14755111Storek 		FPU_SET_CARRY(y->fp_sticky);
14855111Storek 		FPU_SUBCS(r3, x->fp_mant[3], y->fp_mant[3]);
14955111Storek 		FPU_SUBCS(r2, x->fp_mant[2], y->fp_mant[2]);
15055111Storek 		FPU_SUBCS(r1, x->fp_mant[1], y->fp_mant[1]);
15155111Storek 		FPU_SUBC(r0, x->fp_mant[0], y->fp_mant[0]);
15255111Storek 		if (r0 < FP_2) {
15355111Storek 			/* cases i and ii */
15455111Storek 			if ((r0 | r1 | r2 | r3) == 0) {
15555111Storek 				/* case ii */
15655111Storek 				r->fp_class = FPC_ZERO;
15755111Storek 				r->fp_sign = rd == FSR_RD_RM;
15855111Storek 				return (r);
15955111Storek 			}
16055111Storek 		} else {
16155111Storek 			/*
16255111Storek 			 * Oops, case iii.  This can only occur when the
16355111Storek 			 * exponents were equal, in which case neither
16455111Storek 			 * x nor y have sticky bits set.  Flip the sign
16555111Storek 			 * (to y's sign) and negate the result to get y - x.
16655111Storek 			 */
16755111Storek #ifdef DIAGNOSTIC
16855111Storek 			if (x->fp_exp != y->fp_exp || r->fp_sticky)
16955111Storek 				panic("fpu_add");
17055111Storek #endif
17155111Storek 			r->fp_sign = y->fp_sign;
17255111Storek 			FPU_SUBS(r3, 0, r3);
17355111Storek 			FPU_SUBCS(r2, 0, r2);
17455111Storek 			FPU_SUBCS(r1, 0, r1);
17555111Storek 			FPU_SUBC(r0, 0, r0);
17655111Storek 		}
17755111Storek 		r->fp_mant[3] = r3;
17855111Storek 		r->fp_mant[2] = r2;
17955111Storek 		r->fp_mant[1] = r1;
18055111Storek 		r->fp_mant[0] = r0;
18155111Storek 		if (r0 < FP_1)
18255111Storek 			fpu_norm(r);
18355111Storek 	}
18455111Storek 	return (r);
18555111Storek }
186