xref: /freebsd/sys/powerpc/fpu/fpu_emu.h (revision 63f537551380d2dab29fa402ad1269feae17e594)
1 /*	$NetBSD: fpu_emu.h,v 1.3 2005/12/11 12:18:42 christos Exp $ */
2 
3 /*-
4  * SPDX-License-Identifier: BSD-3-Clause
5  *
6  * Copyright (c) 1992, 1993
7  *	The Regents of the University of California.  All rights reserved.
8  *
9  * This software was developed by the Computer Systems Engineering group
10  * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
11  * contributed to Berkeley.
12  *
13  * All advertising materials mentioning features or use of this software
14  * must display the following acknowledgement:
15  *	This product includes software developed by the University of
16  *	California, Lawrence Berkeley Laboratory.
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  * 1. Redistributions of source code must retain the above copyright
22  *    notice, this list of conditions and the following disclaimer.
23  * 2. Redistributions in binary form must reproduce the above copyright
24  *    notice, this list of conditions and the following disclaimer in the
25  *    documentation and/or other materials provided with the distribution.
26  * 3. Neither the name of the University nor the names of its contributors
27  *    may be used to endorse or promote products derived from this software
28  *    without specific prior written permission.
29  *
30  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
31  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
32  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
33  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
34  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
35  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
36  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
37  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
38  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
39  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
40  * SUCH DAMAGE.
41  *
42  *	@(#)fpu_emu.h	8.1 (Berkeley) 6/11/93
43  */
44 
45 /*
46  * Floating point emulator (tailored for SPARC, but structurally
47  * machine-independent).
48  *
49  * Floating point numbers are carried around internally in an `expanded'
50  * or `unpacked' form consisting of:
51  *	- sign
52  *	- unbiased exponent
53  *	- mantissa (`1.' + 112-bit fraction + guard + round)
54  *	- sticky bit
55  * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
56  * always nonzero.  Additional low-order `guard' and `round' bits are
57  * scrunched in, making the entire mantissa 115 bits long.  This is divided
58  * into four 32-bit words, with `spare' bits left over in the upper part
59  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
60  * number is thus kept within the half-open interval [1.0,2.0) (but see
61  * the `number classes' below).  This holds even for denormalized numbers:
62  * when we explode an external denorm, we normalize it, introducing low-order
63  * zero bits, so that the rest of the code always sees normalized values.
64  *
65  * Note that a number of our algorithms use the `spare' bits at the top.
66  * The most demanding algorithm---the one for sqrt---depends on two such
67  * bits, so that it can represent values up to (but not including) 8.0,
68  * and then it needs a carry on top of that, so that we need three `spares'.
69  *
70  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
71  * whole words from the mantissa into it.
72  *
73  * All operations are done in this internal extended precision.  According
74  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
75  * it is OK to do a+b in extended precision and then round the result to
76  * single precision---provided single, double, and extended precisions are
77  * `far enough apart' (they always are), but we will try to avoid any such
78  * extra work where possible.
79  */
80 struct fpn {
81 	int	fp_class;		/* see below */
82 	int	fp_sign;		/* 0 => positive, 1 => negative */
83 	int	fp_exp;			/* exponent (unbiased) */
84 	int	fp_sticky;		/* nonzero bits lost at right end */
85 	u_int	fp_mant[4];		/* 115-bit mantissa */
86 };
87 
88 #define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
89 #define	FP_NG		2		/* number of low-order guard bits */
90 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
91 #define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
92 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
93 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
94 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
95 
96 /*
97  * Number classes.  Since zero, Inf, and NaN cannot be represented using
98  * the above layout, we distinguish these from other numbers via a class.
99  * In addition, to make computation easier and to follow Appendix N of
100  * the SPARC Version 8 standard, we give each kind of NaN a separate class.
101  */
102 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
103 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
104 #define	FPC_ZERO	0		/* zero (sign matters) */
105 #define	FPC_NUM		1		/* number (sign matters) */
106 #define	FPC_INF		2		/* infinity (sign matters) */
107 
108 #define	ISSNAN(fp)	((fp)->fp_class == FPC_SNAN)
109 #define	ISQNAN(fp)	((fp)->fp_class == FPC_QNAN)
110 #define	ISNAN(fp)	((fp)->fp_class < 0)
111 #define	ISZERO(fp)	((fp)->fp_class == 0)
112 #define	ISINF(fp)	((fp)->fp_class == FPC_INF)
113 
114 /*
115  * ORDER(x,y) `sorts' a pair of `fpn *'s so that the right operand (y) points
116  * to the `more significant' operand for our purposes.  Appendix N says that
117  * the result of a computation involving two numbers are:
118  *
119  *	If both are SNaN: operand 2, converted to Quiet
120  *	If only one is SNaN: the SNaN operand, converted to Quiet
121  *	If both are QNaN: operand 2
122  *	If only one is QNaN: the QNaN operand
123  *
124  * In addition, in operations with an Inf operand, the result is usually
125  * Inf.  The class numbers are carefully arranged so that if
126  *	(unsigned)class(op1) > (unsigned)class(op2)
127  * then op1 is the one we want; otherwise op2 is the one we want.
128  */
129 #define	ORDER(x, y) { \
130 	if ((u_int)(x)->fp_class > (u_int)(y)->fp_class) \
131 		SWAP(x, y); \
132 }
133 #define	SWAP(x, y) { \
134 	struct fpn *swap; \
135 	swap = (x), (x) = (y), (y) = swap; \
136 }
137 
138 /*
139  * Emulator state.
140  */
141 struct fpemu {
142 	struct	fpu *fe_fpstate;	/* registers, etc */
143 	int	fe_fpscr;		/* fpscr copy (modified during op) */
144 	int	fe_cx;			/* keep track of exceptions */
145 	struct	fpn fe_f1;		/* operand 1 */
146 	struct	fpn fe_f2;		/* operand 2, if required */
147 	struct	fpn fe_f3;		/* available storage for result */
148 };
149 
150 /*
151  * Arithmetic functions.
152  * Each of these may modify its inputs (f1,f2) and/or the temporary.
153  * Each returns a pointer to the result and/or sets exceptions.
154  */
155 struct	fpn *fpu_add(struct fpemu *);
156 #define	fpu_sub(fe) ((fe)->fe_f2.fp_sign ^= 1, fpu_add(fe))
157 struct	fpn *fpu_mul(struct fpemu *);
158 struct	fpn *fpu_div(struct fpemu *);
159 struct	fpn *fpu_sqrt(struct fpemu *);
160 
161 /*
162  * Other functions.
163  */
164 
165 /* Perform a compare instruction (with or without unordered exception). */
166 void	fpu_compare(struct fpemu *, int);
167 
168 /* Build a new Quiet NaN (sign=0, frac=all 1's). */
169 struct	fpn *fpu_newnan(struct fpemu *);
170 
171 void	fpu_norm(struct fpn *);
172 
173 /*
174  * Shift a number right some number of bits, taking care of round/sticky.
175  * Note that the result is probably not a well-formed number (it will lack
176  * the normal 1-bit mant[0]&FP_1).
177  */
178 int	fpu_shr(struct fpn *, int);
179 
180 void	fpu_explode(struct fpemu *, struct fpn *, int, int);
181 void	fpu_implode(struct fpemu *, struct fpn *, int, u_int *);
182 
183 #ifdef DEBUG
184 #define	FPE_EX		0x1
185 #define	FPE_INSN	0x2
186 #define	FPE_OP		0x4
187 #define	FPE_REG		0x8
188 extern int fpe_debug;
189 void	fpu_dumpfpn(struct fpn *);
190 #define	DPRINTF(x, y)	if (fpe_debug & (x)) printf y
191 #define DUMPFPN(x, f)	if (fpe_debug & (x)) fpu_dumpfpn((f))
192 #else
193 #define	DPRINTF(x, y)
194 #define DUMPFPN(x, f)
195 #endif
196