Home | History | Annotate | Line # | Download | only in fpe
fpu_emulate.h revision 1.6
      1 /*	$NetBSD: fpu_emulate.h,v 1.6 2000/09/19 03:17:59 minoura Exp $	*/
      2 
      3 /*
      4  * Copyright (c) 1995 Gordon Ross
      5  * Copyright (c) 1995 Ken Nakata
      6  * All rights reserved.
      7  *
      8  * Redistribution and use in source and binary forms, with or without
      9  * modification, are permitted provided that the following conditions
     10  * are met:
     11  * 1. Redistributions of source code must retain the above copyright
     12  *    notice, this list of conditions and the following disclaimer.
     13  * 2. Redistributions in binary form must reproduce the above copyright
     14  *    notice, this list of conditions and the following disclaimer in the
     15  *    documentation and/or other materials provided with the distribution.
     16  * 3. The name of the author may not be used to endorse or promote products
     17  *    derived from this software without specific prior written permission.
     18  * 4. All advertising materials mentioning features or use of this software
     19  *    must display the following acknowledgement:
     20  *      This product includes software developed by Gordon Ross
     21  *
     22  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
     23  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
     24  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
     25  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
     26  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
     27  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
     28  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
     29  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
     30  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
     31  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     32  */
     33 
     34 #ifndef _FPU_EMULATE_H_
     35 #define _FPU_EMULATE_H_
     36 
     37 #include <sys/types.h>
     38 
     39 /*
     40  * Floating point emulator (tailored for SPARC/modified for m68k, but
     41  * structurally machine-independent).
     42  *
     43  * Floating point numbers are carried around internally in an `expanded'
     44  * or `unpacked' form consisting of:
     45  *	- sign
     46  *	- unbiased exponent
     47  *	- mantissa (`1.' + 63-bit fraction + guard + round)
     48  *	- sticky bit
     49  * Any implied `1' bit is inserted, giving a 64-bit mantissa that is
     50  * always nonzero.  Additional low-order `guard' and `round' bits are
     51  * scrunched in, making the entire mantissa 66 bits long.  This is divided
     52  * into three 32-bit words, with `spare' bits left over in the upper part
     53  * of the top word (the high bits of fp_mant[0]).  An internal `exploded'
     54  * number is thus kept within the half-open interval [1.0,2.0) (but see
     55  * the `number classes' below).  This holds even for denormalized numbers:
     56  * when we explode an external denorm, we normalize it, introducing low-order
     57  * zero bits, so that the rest of the code always sees normalized values.
     58  *
     59  * Note that a number of our algorithms use the `spare' bits at the top.
     60  * The most demanding algorithm---the one for sqrt---depends on two such
     61  * bits, so that it can represent values up to (but not including) 8.0,
     62  * and then it needs a carry on top of that, so that we need three `spares'.
     63  *
     64  * The sticky-word is 32 bits so that we can use `OR' operators to goosh
     65  * whole words from the mantissa into it.
     66  *
     67  * All operations are done in this internal extended precision.  According
     68  * to Hennesey & Patterson, Appendix A, rounding can be repeated---that is,
     69  * it is OK to do a+b in extended precision and then round the result to
     70  * single precision---provided single, double, and extended precisions are
     71  * `far enough apart' (they always are), but we will try to avoid any such
     72  * extra work where possible.
     73  */
     74 struct fpn {
     75 	int	fp_class;		/* see below */
     76 	int	fp_sign;		/* 0 => positive, 1 => negative */
     77 	int	fp_exp;			/* exponent (unbiased) */
     78 	int	fp_sticky;		/* nonzero bits lost at right end */
     79 	u_int	fp_mant[3];		/* 66-bit mantissa */
     80 };
     81 
     82 #define	FP_NMANT	67		/* total bits in mantissa (incl g,r) */
     83 #define	FP_NG		2		/* number of low-order guard bits */
     84 #define	FP_LG		((FP_NMANT - 1) & 31)	/* log2(1.0) for fp_mant[0] */
     85 #define	FP_QUIETBIT	(1 << (FP_LG - 1))	/* Quiet bit in NaNs (0.5) */
     86 #define	FP_1		(1 << FP_LG)		/* 1.0 in fp_mant[0] */
     87 #define	FP_2		(1 << (FP_LG + 1))	/* 2.0 in fp_mant[0] */
     88 
     89 #define CPYFPN(dst, src)						\
     90 if ((dst) != (src)) {							\
     91     (dst)->fp_class = (src)->fp_class;					\
     92     (dst)->fp_sign = (src)->fp_sign;					\
     93     (dst)->fp_exp = (src)->fp_exp;					\
     94     (dst)->fp_sticky = (src)->fp_sticky;				\
     95     (dst)->fp_mant[0] = (src)->fp_mant[0];				\
     96     (dst)->fp_mant[1] = (src)->fp_mant[1];				\
     97     (dst)->fp_mant[2] = (src)->fp_mant[2];				\
     98 }
     99 
    100 /*
    101  * Number classes.  Since zero, Inf, and NaN cannot be represented using
    102  * the above layout, we distinguish these from other numbers via a class.
    103  */
    104 #define	FPC_SNAN	-2		/* signalling NaN (sign irrelevant) */
    105 #define	FPC_QNAN	-1		/* quiet NaN (sign irrelevant) */
    106 #define	FPC_ZERO	0		/* zero (sign matters) */
    107 #define	FPC_NUM		1		/* number (sign matters) */
    108 #define	FPC_INF		2		/* infinity (sign matters) */
    109 
    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 	register struct fpn *swap;		\
    135 	swap = (x), (x) = (y), (y) = swap;	\
    136 }
    137 
    138 /*
    139  * Emulator state.
    140  */
    141 struct fpemu {
    142     struct	frame *fe_frame; /* integer regs, etc */
    143     struct	fpframe *fe_fpframe; /* FP registers, etc */
    144     u_int	fe_fpsr;	/* fpsr copy (modified during op) */
    145     u_int	fe_fpcr;	/* fpcr copy */
    146     struct	fpn fe_f1;	/* operand 1 */
    147     struct	fpn fe_f2;	/* operand 2, if required */
    148     struct	fpn fe_f3;	/* available storage for result */
    149 };
    150 
    151 /*****************************************************************************
    152  * End of definitions derived from Sparc FPE
    153  *****************************************************************************/
    154 
    155 /*
    156  * Internal info about a decoded effective address.
    157  */
    158 struct insn_ea {
    159     int	ea_regnum;
    160     int	ea_ext[3];		/* extention words if any */
    161     int	ea_flags;		/* flags == 0 means mode 2: An@ */
    162 #define	EA_DIRECT	0x001	/* mode [01]: Dn or An */
    163 #define EA_PREDECR	0x002	/* mode 4: An@- */
    164 #define	EA_POSTINCR	0x004	/* mode 3: An@+ */
    165 #define EA_OFFSET	0x008	/* mode 5 or (7,2): APC@(d16) */
    166 #define	EA_INDEXED	0x010	/* mode 6 or (7,3): APC@(Xn:*:*,d8) etc */
    167 #define EA_ABS  	0x020	/* mode (7,[01]): abs */
    168 #define EA_PC_REL	0x040	/* mode (7,[23]): PC@(d16) etc */
    169 #define	EA_IMMED	0x080	/* mode (7,4): #immed */
    170 #define EA_MEM_INDIR	0x100	/* mode 6 or (7,3): APC@(Xn:*:*,*)@(*) etc */
    171 #define EA_BASE_SUPPRSS	0x200	/* mode 6 or (7,3): base register suppressed */
    172 #define EA_FRAME_EA	0x400	/* MC68LC040 only: precalculated EA from
    173 				   format 4 stack frame */
    174     int	ea_moffs;		/* offset used for fmoveMulti */
    175 };
    176 
    177 #define ea_offset	ea_ext[0]	/* mode 5: offset word */
    178 #define ea_absaddr	ea_ext[0]	/* mode (7,[01]): absolute address */
    179 #define ea_immed	ea_ext		/* mode (7,4): immediate value */
    180 #define ea_basedisp	ea_ext[0]	/* mode 6: base displacement */
    181 #define ea_outerdisp	ea_ext[1]	/* mode 6: outer displacement */
    182 #define	ea_idxreg	ea_ext[2]	/* mode 6: index register number */
    183 #define ea_fea		ea_ext[0]	/* MC68LC040 only: frame EA */
    184 
    185 struct instruction {
    186     u_int		is_pc;		/* insn's address */
    187     u_int		is_nextpc;	/* next PC */
    188     int			is_advance;	/* length of instruction */
    189     int			is_datasize;	/* size of memory operand */
    190     int			is_opcode;	/* opcode word */
    191     int			is_word1;	/* second word */
    192     struct insn_ea	is_ea;	/* decoded effective address mode */
    193 };
    194 
    195 /*
    196  * FP data types
    197  */
    198 #define FTYPE_LNG 0 /* Long Word Integer */
    199 #define FTYPE_SNG 1 /* Single Prec */
    200 #define FTYPE_EXT 2 /* Extended Prec */
    201 #define FTYPE_BCD 3 /* Packed BCD */
    202 #define FTYPE_WRD 4 /* Word Integer */
    203 #define FTYPE_DBL 5 /* Double Prec */
    204 #define FTYPE_BYT 6 /* Byte Integer */
    205 
    206 /*
    207  * MC68881/68882 FPcr bit definitions (should these go to <m68k/reg.h>
    208  * or <m68k/fpu.h> or something?)
    209  */
    210 
    211 /* fpsr */
    212 #define FPSR_CCB    0xff000000
    213 # define FPSR_NEG   0x08000000
    214 # define FPSR_ZERO  0x04000000
    215 # define FPSR_INF   0x02000000
    216 # define FPSR_NAN   0x01000000
    217 #define FPSR_QTT    0x00ff0000
    218 # define FPSR_QSG   0x00800000
    219 # define FPSR_QUO   0x007f0000
    220 #define FPSR_EXCP   0x0000ff00
    221 # define FPSR_BSUN  0x00008000
    222 # define FPSR_SNAN  0x00004000
    223 # define FPSR_OPERR 0x00002000
    224 # define FPSR_OVFL  0x00001000
    225 # define FPSR_UNFL  0x00000800
    226 # define FPSR_DZ    0x00000400
    227 # define FPSR_INEX2 0x00000200
    228 # define FPSR_INEX1 0x00000100
    229 #define FPSR_AEX    0x000000ff
    230 # define FPSR_AIOP  0x00000080
    231 # define FPSR_AOVFL 0x00000040
    232 # define FPSR_AUNFL 0x00000020
    233 # define FPSR_ADZ   0x00000010
    234 # define FPSR_AINEX 0x00000008
    235 
    236 /* fpcr */
    237 #define FPCR_EXCP   FPSR_EXCP
    238 # define FPCR_BSUN  FPSR_BSUN
    239 # define FPCR_SNAN  FPSR_SNAN
    240 # define FPCR_OPERR FPSR_OPERR
    241 # define FPCR_OVFL  FPSR_OVFL
    242 # define FPCR_UNFL  FPSR_UNFL
    243 # define FPCR_DZ    FPSR_DZ
    244 # define FPCR_INEX2 FPSR_INEX2
    245 # define FPCR_INEX1 FPSR_INEX1
    246 #define FPCR_MODE   0x000000ff
    247 # define FPCR_PREC  0x000000c0
    248 #  define FPCR_EXTD 0x00000000
    249 #  define FPCR_SNGL 0x00000040
    250 #  define FPCR_DBL  0x00000080
    251 # define FPCR_ROUND 0x00000030
    252 #  define FPCR_NEAR 0x00000000
    253 #  define FPCR_ZERO 0x00000010
    254 #  define FPCR_MINF 0x00000020
    255 #  define FPCR_PINF 0x00000030
    256 
    257 /*
    258  * Other functions.
    259  */
    260 
    261 /* Build a new Quiet NaN (sign=0, frac=all 1's). */
    262 struct	fpn *fpu_newnan __P((struct fpemu *fe));
    263 
    264 /*
    265  * Shift a number right some number of bits, taking care of round/sticky.
    266  * Note that the result is probably not a well-formed number (it will lack
    267  * the normal 1-bit mant[0]&FP_1).
    268  */
    269 int	fpu_shr __P((struct fpn * fp, int shr));
    270 /*
    271  * Round a number according to the round mode in FPCR
    272  */
    273 int	round __P((register struct fpemu *fe, register struct fpn *fp));
    274 
    275 /* type conversion */
    276 void	fpu_explode __P((struct fpemu *fe, struct fpn *fp, int t, u_int *src));
    277 void	fpu_implode __P((struct fpemu *fe, struct fpn *fp, int t, u_int *dst));
    278 
    279 /*
    280  * non-static emulation functions
    281  */
    282 /* type 0 */
    283 int fpu_emul_fmovecr __P((struct fpemu *fe, struct instruction *insn));
    284 int fpu_emul_fstore __P((struct fpemu *fe, struct instruction *insn));
    285 int fpu_emul_fscale __P((struct fpemu *fe, struct instruction *insn));
    286 
    287 /*
    288  * include function declarations of those which are called by fpu_emul_arith()
    289  */
    290 #include "fpu_arith_proto.h"
    291 
    292 int fpu_emulate __P((struct frame *frame, struct fpframe *fpf));
    293 
    294 /*
    295  * "helper" functions
    296  */
    297 /* return values from constant rom */
    298 struct fpn *fpu_const __P((struct fpn *fp, u_int offset));
    299 /* update exceptions and FPSR */
    300 int fpu_upd_excp __P((struct fpemu *fe));
    301 u_int fpu_upd_fpsr __P((struct fpemu *fe, struct fpn *fp));
    302 
    303 /* address mode decoder, and load/store */
    304 int fpu_decode_ea __P((struct frame *frame, struct instruction *insn,
    305 		   struct insn_ea *ea, int modreg));
    306 int fpu_load_ea __P((struct frame *frame, struct instruction *insn,
    307 		 struct insn_ea *ea, char *dst));
    308 int fpu_store_ea __P((struct frame *frame, struct instruction *insn,
    309 		  struct insn_ea *ea, char *src));
    310 
    311 /* fpu_subr.c */
    312 void fpu_norm __P((register struct fpn *fp));
    313 
    314 #if !defined(FPE_DEBUG)
    315 #  define FPE_DEBUG 0
    316 #endif
    317 
    318 #endif /* _FPU_EMULATE_H_ */
    319