xref: /netbsd-src/sys/arch/powerpc/fpu/fpu_emu.h (revision 89b83c1832f88faa9fda2dbf5e415a8b202c5a46)
1 /*	$NetBSD: fpu_emu.h,v 1.12 2022/09/07 02:41:39 rin 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. Neither the name of the University nor the names of its contributors
25  *    may be used to endorse or promote products derived from this software
26  *    without specific prior written permission.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
29  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
30  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
31  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
32  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
33  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
34  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
35  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
36  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
37  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38  * SUCH DAMAGE.
39  *
40  *	@(#)fpu_emu.h	8.1 (Berkeley) 6/11/93
41  */
42 
43 /*
44  * Floating point emulator (tailored for SPARC, but structurally
45  * machine-independent).
46  *
47  * Floating point numbers are carried around internally in an `expanded'
48  * or `unpacked' form consisting of:
49  *	- sign
50  *	- unbiased exponent
51  *	- mantissa (`1.' + 112-bit fraction + guard + round)
52  *	- sticky bit
53  * Any implied `1' bit is inserted, giving a 113-bit mantissa that is
54  * always nonzero.  Additional low-order `guard' and `round' bits are
55  * scrunched in, making the entire mantissa 115 bits long.  This is divided
56  * into four 32-bit words, with `spare' bits left over in the upper part
57  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
58  * number is thus kept within the half-open interval [1.0,2.0) (but see
59  * the `number classes' below).  This holds even for denormalized numbers:
60  * when we explode an external denorm, we normalize it, introducing low-order
61  * zero bits, so that the rest of the code always sees normalized values.
62  *
63  * Note that a number of our algorithms use the `spare' bits at the top.
64  * The most demanding algorithm---the one for sqrt---depends on two such
65  * bits, so that it can represent values up to (but not including) 8.0,
66  * and then it needs a carry on top of that, so that we need three `spares'.
67  *
68  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
69  * whole words from the mantissa into it.
70  *
71  * All operations are done in this internal extended precision.  According
72  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
73  * it is OK to do a+b in extended precision and then round the result to
74  * single precision---provided single, double, and extended precisions are
75  * `far enough apart' (they always are), but we will try to avoid any such
76  * extra work where possible.
77  */
78 struct fpn {
79 	int	fp_class;		/* see below */
80 	int	fp_sign;		/* 0 => positive, 1 => negative */
81 	int	fp_exp;			/* exponent (unbiased) */
82 	int	fp_sticky;		/* nonzero bits lost at right end */
83 	u_int	fp_mant[4];		/* 115-bit mantissa */
84 };
85 
86 #define	FP_NMANT	115		/* total bits in mantissa (incl g,r) */
87 #define	FP_NG		2		/* number of low-order guard bits */
88 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
89 #define	FP_LG2		((FP_NMANT - 1) & 63)	/* log2(1.0) for fp_mant[0] and fp_mant[1] */
90 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
91 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
92 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
93 
94 /*
95  * Number classes.  Since zero, Inf, and NaN cannot be represented using
96  * the above layout, we distinguish these from other numbers via a class.
97  * In addition, to make computation easier and to follow Appendix N of
98  * the SPARC Version 8 standard, we give each kind of NaN a separate class.
99  */
100 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
101 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
102 #define	FPC_ZERO	0		/* zero (sign matters) */
103 #define	FPC_NUM		1		/* number (sign matters) */
104 #define	FPC_INF		2		/* infinity (sign matters) */
105 
106 #define	ISSNAN(fp)	((fp)->fp_class == FPC_SNAN)
107 #define	ISQNAN(fp)	((fp)->fp_class == FPC_QNAN)
108 #define	ISNAN(fp)	((fp)->fp_class < 0)
109 #define	ISZERO(fp)	((fp)->fp_class == 0)
110 #define	ISINF(fp)	((fp)->fp_class == FPC_INF)
111 
112 /*
113  * ORDER(x,y) `sorts' a pair of `fpn *'s so that the right operand (y) points
114  * to the `more significant' operand for our purposes.  Appendix N says that
115  * the result of a computation involving two numbers are:
116  *
117  *	If both are SNaN: operand 2, converted to Quiet
118  *	If only one is SNaN: the SNaN operand, converted to Quiet
119  *	If both are QNaN: operand 2
120  *	If only one is QNaN: the QNaN operand
121  *
122  * In addition, in operations with an Inf operand, the result is usually
123  * Inf.  The class numbers are carefully arranged so that if
124  *	(unsigned)class(op1) > (unsigned)class(op2)
125  * then op1 is the one we want; otherwise op2 is the one we want.
126  */
127 #define	ORDER(x, y) { \
128 	if ((u_int)(x)->fp_class > (u_int)(y)->fp_class) \
129 		SWAP(x, y); \
130 }
131 #define	SWAP(x, y) { \
132 	struct fpn *swap; \
133 	swap = (x), (x) = (y), (y) = swap; \
134 }
135 
136 /*
137  * FPU data types.
138  */
139 #define	FTYPE_INT	0x00	/* data = 32-bit signed integer */
140 #define	FTYPE_LNG	0x01	/* data = 64-bit signed long integer */
141 #define	FTYPE_SNG	0x02	/* data = 32-bit float */
142 #define	FTYPE_DBL	0x04	/* data = 64-bit double */
143 #define	FTYPE_RD_RZ	0x08
144 #define	FTYPE_FPSCR	0x10
145 #define	FTYPE_FLAG_MASK	(FTYPE_RD_RZ | FTYPE_FPSCR)
146 
147 /*
148  * Emulator state.
149  */
150 struct fpemu {
151 	struct	fpreg *fe_fpstate;	/* registers, etc */
152 	int	fe_fpscr;		/* fpscr copy (modified during op) */
153 	int	fe_cx;			/* keep track of exceptions */
154 	struct	fpn fe_f1;		/* operand 1 */
155 	struct	fpn fe_f2;		/* operand 2, if required */
156 	struct	fpn fe_f3;		/* available storage for result */
157 	vaddr_t fe_addr;		/* last address accessed */
158 };
159 
160 /*
161  * Arithmetic functions.
162  * Each of these may modify its inputs (f1,f2) and/or the temporary.
163  * Each returns a pointer to the result and/or sets exceptions.
164  */
165 struct	fpn *fpu_add(struct fpemu *);
166 struct	fpn *fpu_mul(struct fpemu *);
167 struct	fpn *fpu_div(struct fpemu *);
168 struct	fpn *fpu_sqrt(struct fpemu *);
169 
170 static inline struct fpn *
fpu_sub(struct fpemu * fe)171 fpu_sub(struct fpemu *fe)
172 {
173 	struct fpn *fp = &fe->fe_f2;
174 
175 	if (!ISNAN(fp))
176 		fp->fp_sign ^= 1;
177 	return fpu_add(fe);
178 }
179 
180 /*
181  * Other functions.
182  */
183 
184 /* Perform a compare instruction (with or without unordered exception). */
185 void	fpu_compare(struct fpemu *, int);
186 
187 /* Build a new Quiet NaN (sign=0, frac=all 1's). */
188 struct	fpn *fpu_newnan(struct fpemu *);
189 
190 /*
191  * Shift a number right some number of bits, taking care of round/sticky.
192  * Note that the result is probably not a well-formed number (it will lack
193  * the normal 1-bit mant[0]&FP_1).
194  */
195 int	fpu_shr(struct fpn *, int);
196 
197 void	fpu_norm(struct fpn *);
198 
199 void	fpu_explode(struct fpemu *, struct fpn *, int, uint64_t);
200 void	fpu_implode(struct fpemu *, struct fpn *, int, uint64_t *);
201 
202 #ifdef DEBUG
203 #define	FPE_EX		0x1
204 #define	FPE_INSN	0x2
205 #define	FPE_OP		0x4
206 #define	FPE_REG		0x8
207 extern int fpe_debug;
208 void	fpu_dumpfpn(struct fpn *);
209 #define	DPRINTF(x, y)	if (fpe_debug & (x)) printf y
210 #define DUMPFPN(x, f)	if (fpe_debug & (x)) fpu_dumpfpn((f))
211 #else
212 #define	DPRINTF(x, y)
213 #define DUMPFPN(x, f)
214 #endif
215