xref: /freebsd/sys/powerpc/fpu/fpu_emu.h (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 /*	$NetBSD: fpu_emu.h,v 1.3 2005/12/11 12:18:42 christos Exp $ */
2 /* $FreeBSD$ */
3 
4 /*-
5  * SPDX-License-Identifier: BSD-3-Clause
6  *
7  * Copyright (c) 1992, 1993
8  *	The Regents of the University of California.  All rights reserved.
9  *
10  * This software was developed by the Computer Systems Engineering group
11  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
12  * contributed to Berkeley.
13  *
14  * All advertising materials mentioning features or use of this software
15  * must display the following acknowledgement:
16  *	This product includes software developed by the University of
17  *	California, Lawrence Berkeley Laboratory.
18  *
19  * Redistribution and use in source and binary forms, with or without
20  * modification, are permitted provided that the following conditions
21  * are met:
22  * 1. Redistributions of source code must retain the above copyright
23  *    notice, this list of conditions and the following disclaimer.
24  * 2. Redistributions in binary form must reproduce the above copyright
25  *    notice, this list of conditions and the following disclaimer in the
26  *    documentation and/or other materials provided with the distribution.
27  * 3. Neither the name of the University nor the names of its contributors
28  *    may be used to endorse or promote products derived from this software
29  *    without specific prior written permission.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41  * SUCH DAMAGE.
42  *
43  *	@(#)fpu_emu.h	8.1 (Berkeley) 6/11/93
44  */
45 
46 /*
47  * Floating point emulator (tailored for SPARC, but structurally
48  * machine-independent).
49  *
50  * Floating point numbers are carried around internally in an `expanded'
51  * or `unpacked' form consisting of:
52  *	- sign
53  *	- unbiased exponent
54  *	- mantissa (`1.' + 112-bit fraction + guard + round)
55  *	- sticky bit
56  * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
57  * always nonzero.  Additional low-order `guard' and `round' bits are
58  * scrunched in, making the entire mantissa 115 bits long.  This is divided
59  * into four 32-bit words, with `spare' bits left over in the upper part
60  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
61  * number is thus kept within the half-open interval [1.0,2.0) (but see
62  * the `number classes' below).  This holds even for denormalized numbers:
63  * when we explode an external denorm, we normalize it, introducing low-order
64  * zero bits, so that the rest of the code always sees normalized values.
65  *
66  * Note that a number of our algorithms use the `spare' bits at the top.
67  * The most demanding algorithm---the one for sqrt---depends on two such
68  * bits, so that it can represent values up to (but not including) 8.0,
69  * and then it needs a carry on top of that, so that we need three `spares'.
70  *
71  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
72  * whole words from the mantissa into it.
73  *
74  * All operations are done in this internal extended precision.  According
75  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
76  * it is OK to do a+b in extended precision and then round the result to
77  * single precision---provided single, double, and extended precisions are
78  * `far enough apart' (they always are), but we will try to avoid any such
79  * extra work where possible.
80  */
81 struct fpn {
82 	int	fp_class;		/* see below */
83 	int	fp_sign;		/* 0 => positive, 1 => negative */
84 	int	fp_exp;			/* exponent (unbiased) */
85 	int	fp_sticky;		/* nonzero bits lost at right end */
86 	u_int	fp_mant[4];		/* 115-bit mantissa */
87 };
88 
89 #define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
90 #define	FP_NG		2		/* number of low-order guard bits */
91 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
92 #define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
93 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
94 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
95 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
96 
97 /*
98  * Number classes.  Since zero, Inf, and NaN cannot be represented using
99  * the above layout, we distinguish these from other numbers via a class.
100  * In addition, to make computation easier and to follow Appendix N of
101  * the SPARC Version 8 standard, we give each kind of NaN a separate class.
102  */
103 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
104 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
105 #define	FPC_ZERO	0		/* zero (sign matters) */
106 #define	FPC_NUM		1		/* number (sign matters) */
107 #define	FPC_INF		2		/* infinity (sign matters) */
108 
109 #define	ISSNAN(fp)	((fp)->fp_class == FPC_SNAN)
110 #define	ISQNAN(fp)	((fp)->fp_class == FPC_QNAN)
111 #define	ISNAN(fp)	((fp)->fp_class < 0)
112 #define	ISZERO(fp)	((fp)->fp_class == 0)
113 #define	ISINF(fp)	((fp)->fp_class == FPC_INF)
114 
115 /*
116  * ORDER(x,y) `sorts' a pair of `fpn *'s so that the right operand (y) points
117  * to the `more significant' operand for our purposes.  Appendix N says that
118  * the result of a computation involving two numbers are:
119  *
120  *	If both are SNaN: operand 2, converted to Quiet
121  *	If only one is SNaN: the SNaN operand, converted to Quiet
122  *	If both are QNaN: operand 2
123  *	If only one is QNaN: the QNaN operand
124  *
125  * In addition, in operations with an Inf operand, the result is usually
126  * Inf.  The class numbers are carefully arranged so that if
127  *	(unsigned)class(op1) > (unsigned)class(op2)
128  * then op1 is the one we want; otherwise op2 is the one we want.
129  */
130 #define	ORDER(x, y) { \
131 	if ((u_int)(x)->fp_class > (u_int)(y)->fp_class) \
132 		SWAP(x, y); \
133 }
134 #define	SWAP(x, y) { \
135 	struct fpn *swap; \
136 	swap = (x), (x) = (y), (y) = swap; \
137 }
138 
139 /*
140  * Emulator state.
141  */
142 struct fpemu {
143 	struct	fpu *fe_fpstate;	/* registers, etc */
144 	int	fe_fpscr;		/* fpscr copy (modified during op) */
145 	int	fe_cx;			/* keep track of exceptions */
146 	struct	fpn fe_f1;		/* operand 1 */
147 	struct	fpn fe_f2;		/* operand 2, if required */
148 	struct	fpn fe_f3;		/* available storage for result */
149 };
150 
151 /*
152  * Arithmetic functions.
153  * Each of these may modify its inputs (f1,f2) and/or the temporary.
154  * Each returns a pointer to the result and/or sets exceptions.
155  */
156 struct	fpn *fpu_add(struct fpemu *);
157 #define	fpu_sub(fe) ((fe)->fe_f2.fp_sign ^= 1, fpu_add(fe))
158 struct	fpn *fpu_mul(struct fpemu *);
159 struct	fpn *fpu_div(struct fpemu *);
160 struct	fpn *fpu_sqrt(struct fpemu *);
161 
162 /*
163  * Other functions.
164  */
165 
166 /* Perform a compare instruction (with or without unordered exception). */
167 void	fpu_compare(struct fpemu *, int);
168 
169 /* Build a new Quiet NaN (sign=0, frac=all 1's). */
170 struct	fpn *fpu_newnan(struct fpemu *);
171 
172 void	fpu_norm(struct fpn *);
173 
174 /*
175  * Shift a number right some number of bits, taking care of round/sticky.
176  * Note that the result is probably not a well-formed number (it will lack
177  * the normal 1-bit mant[0]&FP_1).
178  */
179 int	fpu_shr(struct fpn *, int);
180 
181 void	fpu_explode(struct fpemu *, struct fpn *, int, int);
182 void	fpu_implode(struct fpemu *, struct fpn *, int, u_int *);
183 
184 #ifdef DEBUG
185 #define	FPE_EX		0x1
186 #define	FPE_INSN	0x2
187 #define	FPE_OP		0x4
188 #define	FPE_REG		0x8
189 extern int fpe_debug;
190 void	fpu_dumpfpn(struct fpn *);
191 #define	DPRINTF(x, y)	if (fpe_debug & (x)) printf y
192 #define DUMPFPN(x, f)	if (fpe_debug & (x)) fpu_dumpfpn((f))
193 #else
194 #define	DPRINTF(x, y)
195 #define DUMPFPN(x, f)
196 #endif
197