xref: /freebsd/sys/powerpc/fpu/fpu_emu.h (revision 7447ca0eb235974642312b9555caec00b57d8fc1)
1 /*	$NetBSD: fpu_emu.h,v 1.3 2005/12/11 12:18:42 christos Exp $ */
2 /* $FreeBSD$ */
3 
4 /*
5  * Copyright (c) 1992, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  *
8  * This software was developed by the Computer Systems Engineering group
9  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
10  * contributed to Berkeley.
11  *
12  * All advertising materials mentioning features or use of this software
13  * must display the following acknowledgement:
14  *	This product includes software developed by the University of
15  *	California, Lawrence Berkeley Laboratory.
16  *
17  * Redistribution and use in source and binary forms, with or without
18  * modification, are permitted provided that the following conditions
19  * are met:
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  * 3. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)fpu_emu.h	8.1 (Berkeley) 6/11/93
42  */
43 
44 /*
45  * Floating point emulator (tailored for SPARC, but structurally
46  * machine-independent).
47  *
48  * Floating point numbers are carried around internally in an `expanded'
49  * or `unpacked' form consisting of:
50  *	- sign
51  *	- unbiased exponent
52  *	- mantissa (`1.' + 112-bit fraction + guard + round)
53  *	- sticky bit
54  * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
55  * always nonzero.  Additional low-order `guard' and `round' bits are
56  * scrunched in, making the entire mantissa 115 bits long.  This is divided
57  * into four 32-bit words, with `spare' bits left over in the upper part
58  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
59  * number is thus kept within the half-open interval [1.0,2.0) (but see
60  * the `number classes' below).  This holds even for denormalized numbers:
61  * when we explode an external denorm, we normalize it, introducing low-order
62  * zero bits, so that the rest of the code always sees normalized values.
63  *
64  * Note that a number of our algorithms use the `spare' bits at the top.
65  * The most demanding algorithm---the one for sqrt---depends on two such
66  * bits, so that it can represent values up to (but not including) 8.0,
67  * and then it needs a carry on top of that, so that we need three `spares'.
68  *
69  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
70  * whole words from the mantissa into it.
71  *
72  * All operations are done in this internal extended precision.  According
73  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
74  * it is OK to do a+b in extended precision and then round the result to
75  * single precision---provided single, double, and extended precisions are
76  * `far enough apart' (they always are), but we will try to avoid any such
77  * extra work where possible.
78  */
79 struct fpn {
80 	int	fp_class;		/* see below */
81 	int	fp_sign;		/* 0 => positive, 1 => negative */
82 	int	fp_exp;			/* exponent (unbiased) */
83 	int	fp_sticky;		/* nonzero bits lost at right end */
84 	u_int	fp_mant[4];		/* 115-bit mantissa */
85 };
86 
87 #define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
88 #define	FP_NG		2		/* number of low-order guard bits */
89 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
90 #define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
91 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
92 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
93 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
94 
95 /*
96  * Number classes.  Since zero, Inf, and NaN cannot be represented using
97  * the above layout, we distinguish these from other numbers via a class.
98  * In addition, to make computation easier and to follow Appendix N of
99  * the SPARC Version 8 standard, we give each kind of NaN a separate class.
100  */
101 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
102 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
103 #define	FPC_ZERO	0		/* zero (sign matters) */
104 #define	FPC_NUM		1		/* number (sign matters) */
105 #define	FPC_INF		2		/* infinity (sign matters) */
106 
107 #define	ISSNAN(fp)	((fp)->fp_class == FPC_SNAN)
108 #define	ISQNAN(fp)	((fp)->fp_class == FPC_QNAN)
109 #define	ISNAN(fp)	((fp)->fp_class < 0)
110 #define	ISZERO(fp)	((fp)->fp_class == 0)
111 #define	ISINF(fp)	((fp)->fp_class == FPC_INF)
112 
113 /*
114  * ORDER(x,y) `sorts' a pair of `fpn *'s so that the right operand (y) points
115  * to the `more significant' operand for our purposes.  Appendix N says that
116  * the result of a computation involving two numbers are:
117  *
118  *	If both are SNaN: operand 2, converted to Quiet
119  *	If only one is SNaN: the SNaN operand, converted to Quiet
120  *	If both are QNaN: operand 2
121  *	If only one is QNaN: the QNaN operand
122  *
123  * In addition, in operations with an Inf operand, the result is usually
124  * Inf.  The class numbers are carefully arranged so that if
125  *	(unsigned)class(op1) > (unsigned)class(op2)
126  * then op1 is the one we want; otherwise op2 is the one we want.
127  */
128 #define	ORDER(x, y) { \
129 	if ((u_int)(x)->fp_class > (u_int)(y)->fp_class) \
130 		SWAP(x, y); \
131 }
132 #define	SWAP(x, y) { \
133 	struct fpn *swap; \
134 	swap = (x), (x) = (y), (y) = swap; \
135 }
136 
137 /*
138  * Emulator state.
139  */
140 struct fpemu {
141 	struct	fpu *fe_fpstate;	/* registers, etc */
142 	int	fe_fpscr;		/* fpscr copy (modified during op) */
143 	int	fe_cx;			/* keep track of exceptions */
144 	struct	fpn fe_f1;		/* operand 1 */
145 	struct	fpn fe_f2;		/* operand 2, if required */
146 	struct	fpn fe_f3;		/* available storage for result */
147 };
148 
149 /*
150  * Arithmetic functions.
151  * Each of these may modify its inputs (f1,f2) and/or the temporary.
152  * Each returns a pointer to the result and/or sets exceptions.
153  */
154 struct	fpn *fpu_add(struct fpemu *);
155 #define	fpu_sub(fe) ((fe)->fe_f2.fp_sign ^= 1, fpu_add(fe))
156 struct	fpn *fpu_mul(struct fpemu *);
157 struct	fpn *fpu_div(struct fpemu *);
158 struct	fpn *fpu_sqrt(struct fpemu *);
159 
160 /*
161  * Other functions.
162  */
163 
164 /* Perform a compare instruction (with or without unordered exception). */
165 void	fpu_compare(struct fpemu *, int);
166 
167 /* Build a new Quiet NaN (sign=0, frac=all 1's). */
168 struct	fpn *fpu_newnan(struct fpemu *);
169 
170 void	fpu_norm(struct fpn *);
171 
172 /*
173  * Shift a number right some number of bits, taking care of round/sticky.
174  * Note that the result is probably not a well-formed number (it will lack
175  * the normal 1-bit mant[0]&FP_1).
176  */
177 int	fpu_shr(struct fpn *, int);
178 
179 void	fpu_explode(struct fpemu *, struct fpn *, int, int);
180 void	fpu_implode(struct fpemu *, struct fpn *, int, u_int *);
181 
182 #ifdef DEBUG
183 #define	FPE_EX		0x1
184 #define	FPE_INSN	0x2
185 #define	FPE_OP		0x4
186 #define	FPE_REG		0x8
187 extern int fpe_debug;
188 void	fpu_dumpfpn(struct fpn *);
189 #define	DPRINTF(x, y)	if (fpe_debug & (x)) printf y
190 #define DUMPFPN(x, f)	if (fpe_debug & (x)) fpu_dumpfpn((f))
191 #else
192 #define	DPRINTF(x, y)
193 #define DUMPFPN(x, f)
194 #endif
195