xref: /openbsd/lib/libc/arch/sparc64/fpu/fpu_emu.h (revision 3cab2bb3)
1 /*	$OpenBSD: fpu_emu.h,v 1.1 2003/07/21 18:41:30 jason 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_emu.h	8.1 (Berkeley) 6/11/93
45  *	$NetBSD: fpu_emu.h,v 1.4 2000/08/03 18:32:07 eeh Exp $
46  * $FreeBSD: src/lib/libc/sparc64/fpu/fpu_emu.h,v 1.4 2002/03/22 23:41:59 obrien Exp $
47  */
48 
49 /*
50  * Floating point emulator (tailored for SPARC, but structurally
51  * machine-independent).
52  *
53  * Floating point numbers are carried around internally in an `expanded'
54  * or `unpacked' form consisting of:
55  *	- sign
56  *	- unbiased exponent
57  *	- mantissa (`1.' + 112-bit fraction + guard + round)
58  *	- sticky bit
59  * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
60  * always nonzero.  Additional low-order `guard' and `round' bits are
61  * scrunched in, making the entire mantissa 115 bits long.  This is divided
62  * into four 32-bit words, with `spare' bits left over in the upper part
63  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
64  * number is thus kept within the half-open interval [1.0,2.0) (but see
65  * the `number classes' below).  This holds even for denormalized numbers:
66  * when we explode an external denorm, we normalize it, introducing low-order
67  * zero bits, so that the rest of the code always sees normalized values.
68  *
69  * Note that a number of our algorithms use the `spare' bits at the top.
70  * The most demanding algorithm---the one for sqrt---depends on two such
71  * bits, so that it can represent values up to (but not including) 8.0,
72  * and then it needs a carry on top of that, so that we need three `spares'.
73  *
74  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
75  * whole words from the mantissa into it.
76  *
77  * All operations are done in this internal extended precision.  According
78  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
79  * it is OK to do a+b in extended precision and then round the result to
80  * single precision---provided single, double, and extended precisions are
81  * `far enough apart' (they always are), but we will try to avoid any such
82  * extra work where possible.
83  */
84 
85 #ifndef _SPARC64_FPU_FPU_EMU_H_
86 #define _SPARC64_FPU_FPU_EMU_H_
87 
88 #include "fpu_reg.h"
89 
90 struct fpn {
91 	int	fp_class;		/* see below */
92 	int	fp_sign;		/* 0 => positive, 1 => negative */
93 	int	fp_exp;			/* exponent (unbiased) */
94 	int	fp_sticky;		/* nonzero bits lost at right end */
95 	u_int	fp_mant[4];		/* 115-bit mantissa */
96 };
97 
98 #define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
99 #define	FP_NG		2		/* number of low-order guard bits */
100 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
101 #define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
102 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
103 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
104 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
105 
106 /*
107  * Number classes.  Since zero, Inf, and NaN cannot be represented using
108  * the above layout, we distinguish these from other numbers via a class.
109  * In addition, to make computation easier and to follow Appendix N of
110  * the SPARC Version 8 standard, we give each kind of NaN a separate class.
111  */
112 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
113 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
114 #define	FPC_ZERO	0		/* zero (sign matters) */
115 #define	FPC_NUM		1		/* number (sign matters) */
116 #define	FPC_INF		2		/* infinity (sign matters) */
117 
118 #define	ISNAN(fp)	((fp)->fp_class < 0)
119 #define	ISZERO(fp)	((fp)->fp_class == 0)
120 #define	ISINF(fp)	((fp)->fp_class == FPC_INF)
121 
122 /*
123  * ORDER(x,y) `sorts' a pair of `fpn *'s so that the right operand (y) points
124  * to the `more significant' operand for our purposes.  Appendix N says that
125  * the result of a computation involving two numbers are:
126  *
127  *	If both are SNaN: operand 2, converted to Quiet
128  *	If only one is SNaN: the SNaN operand, converted to Quiet
129  *	If both are QNaN: operand 2
130  *	If only one is QNaN: the QNaN operand
131  *
132  * In addition, in operations with an Inf operand, the result is usually
133  * Inf.  The class numbers are carefully arranged so that if
134  *	(unsigned)class(op1) > (unsigned)class(op2)
135  * then op1 is the one we want; otherwise op2 is the one we want.
136  */
137 #define	ORDER(x, y) { \
138 	if ((u_int)(x)->fp_class > (u_int)(y)->fp_class) \
139 		SWAP(x, y); \
140 }
141 #define	SWAP(x, y) { \
142 	register struct fpn *swap; \
143 	swap = (x), (x) = (y), (y) = swap; \
144 }
145 
146 /*
147  * Emulator state.
148  */
149 struct fpemu {
150 	u_long	fe_fsr;			/* fsr copy (modified during op) */
151 	int	fe_cx;			/* exceptions */
152 	struct	fpn fe_f1;		/* operand 1 */
153 	struct	fpn fe_f2;		/* operand 2, if required */
154 	struct	fpn fe_f3;		/* available storage for result */
155 };
156 
157 /*
158  * Arithmetic functions.
159  * Each of these may modify its inputs (f1,f2) and/or the temporary.
160  * Each returns a pointer to the result and/or sets exceptions.
161  */
162 #define	__fpu_sub(fe) ((fe)->fe_f2.fp_sign ^= 1, __fpu_add(fe))
163 
164 #ifdef FPU_DEBUG
165 #define	FPE_INSN	0x1
166 #define	FPE_REG		0x2
167 extern int __fpe_debug;
168 void	__fpu_dumpfpn(struct fpn *);
169 #define	DPRINTF(x, y)	if (__fpe_debug & (x)) printf y
170 #define DUMPFPN(x, f)	if (__fpe_debug & (x)) __fpu_dumpfpn((f))
171 #else
172 #define	DPRINTF(x, y)
173 #define DUMPFPN(x, f)
174 #endif
175 
176 #define FSR_GET_RD(fsr)		(((fsr) >> FSR_RD_SHIFT) & FSR_RD_MASK)
177 #endif /* !_SPARC64_FPU_FPU_EXTERN_H_ */
178