1/*-
2 * SPDX-License-Identifier: BSD-3-Clause
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 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 *    notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 *    notice, this list of conditions and the following disclaimer in the
18 *    documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 *    may be used to endorse or promote products derived from this software
21 *    without specific prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 *
35 *	@(#)fpu_emu.h	8.1 (Berkeley) 6/11/93
36 *	$NetBSD: fpu_emu.h,v 1.4 2000/08/03 18:32:07 eeh Exp $
37 * $FreeBSD: head/lib/libc/sparc64/fpu/fpu_emu.h 326025 2017-11-20 19:49:47Z pfg $
38 */
39
40/*
41 * Floating point emulator (tailored for SPARC, but structurally
42 * machine-independent).
43 *
44 * Floating point numbers are carried around internally in an `expanded'
45 * or `unpacked' form consisting of:
46 *	- sign
47 *	- unbiased exponent
48 *	- mantissa (`1.' + 112-bit fraction + guard + round)
49 *	- sticky bit
50 * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
51 * always nonzero.  Additional low-order `guard' and `round' bits are
52 * scrunched in, making the entire mantissa 115 bits long.  This is divided
53 * into four 32-bit words, with `spare' bits left over in the upper part
54 * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
55 * number is thus kept within the half-open interval [1.0,2.0) (but see
56 * the `number classes' below).  This holds even for denormalized numbers:
57 * when we explode an external denorm, we normalize it, introducing low-order
58 * zero bits, so that the rest of the code always sees normalized values.
59 *
60 * Note that a number of our algorithms use the `spare' bits at the top.
61 * The most demanding algorithm---the one for sqrt---depends on two such
62 * bits, so that it can represent values up to (but not including) 8.0,
63 * and then it needs a carry on top of that, so that we need three `spares'.
64 *
65 * The sticky-word is 32 bits so that we can use `OR' operators to goosh
66 * whole words from the mantissa into it.
67 *
68 * All operations are done in this internal extended precision.  According
69 * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
70 * it is OK to do a+b in extended precision and then round the result to
71 * single precision---provided single, double, and extended precisions are
72 * `far enough apart' (they always are), but we will try to avoid any such
73 * extra work where possible.
74 */
75
76#ifndef _SPARC64_FPU_FPU_EMU_H_
77#define _SPARC64_FPU_FPU_EMU_H_
78
79#include <stdint.h>
80
81#include "fpu_reg.h"
82
83struct fpn {
84	int	fp_class;		/* see below */
85	int	fp_sign;		/* 0 => positive, 1 => negative */
86	int	fp_exp;			/* exponent (unbiased) */
87	int	fp_sticky;		/* nonzero bits lost at right end */
88	uint32_t	fp_mant[4];		/* 115-bit mantissa */
89};
90
91#define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
92#define	FP_NG		2		/* number of low-order guard bits */
93#define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
94#define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
95#define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
96#define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
97#define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
98
99/*
100 * Number classes.  Since zero, Inf, and NaN cannot be represented using
101 * the above layout, we distinguish these from other numbers via a class.
102 * In addition, to make computation easier and to follow Appendix N of
103 * the SPARC Version 8 standard, we give each kind of NaN a separate class.
104 */
105#define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
106#define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
107#define	FPC_ZERO	0		/* zero (sign matters) */
108#define	FPC_NUM		1		/* number (sign matters) */
109#define	FPC_INF		2		/* infinity (sign matters) */
110
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 ((uint32_t)(x)->fp_class > (uint32_t)(y)->fp_class) \
132		SWAP(x, y); \
133}
134#define	SWAP(x, y) { \
135	register struct fpn *swap; \
136	swap = (x), (x) = (y), (y) = swap; \
137}
138
139/*
140 * Floating point operand types. FTYPE_LNG is syntethic (it does not occur in
141 * instructions).
142 */
143#define	FTYPE_INT	INSFP_i
144#define	FTYPE_SNG	INSFP_s
145#define	FTYPE_DBL	INSFP_d
146#define	FTYPE_EXT	INSFP_q
147#define	FTYPE_LNG	4
148
149/*
150 * Emulator state.
151 */
152struct fpemu {
153	uint64_t	fe_fsr;			/* fsr copy (modified during op) */
154	int	fe_cx;			/* exceptions */
155	int     pad;                    /* align access to following fields */
156	struct	fpn fe_f1;		/* operand 1 */
157	struct	fpn fe_f2;		/* operand 2, if required */
158	struct	fpn fe_f3;		/* available storage for result */
159};
160
161/*
162 * Arithmetic functions.
163 * Each of these may modify its inputs (f1,f2) and/or the temporary.
164 * Each returns a pointer to the result and/or sets exceptions.
165 */
166#define	__fpu_sub(fe)	(ISNAN(&(fe)->fe_f2) ? 0 : ((fe)->fe_f2.fp_sign ^= 1), \
167			    __fpu_add(fe))
168
169#ifdef FPU_DEBUG
170#define	FPE_INSN	0x1
171#define	FPE_REG		0x2
172extern int __fpe_debug;
173void	__fpu_dumpfpn(struct fpn *);
174#define	DPRINTF(x, y)	if (__fpe_debug & (x)) printf y
175#define DUMPFPN(x, f)	if (__fpe_debug & (x)) __fpu_dumpfpn((f))
176#else
177#define	DPRINTF(x, y)
178#define DUMPFPN(x, f)
179#endif
180
181#endif /* !_SPARC64_FPU_FPU_EXTERN_H_ */
182