1/* $NetBSD: fpu_emu.h,v 1.1.24.3 2004/09/21 13:20:34 skrll 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 */ 78struct 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 * Emulator state. 138 */ 139struct fpemu { 140 struct fpreg *fe_fpstate; /* registers, etc */ 141 int fe_fpscr; /* fpscr copy (modified during op) */ 142 int fe_cx; /* keep track of exceptions */ 143 struct fpn fe_f1; /* operand 1 */ 144 struct fpn fe_f2; /* operand 2, if required */ 145 struct fpn fe_f3; /* available storage for result */ 146}; 147 148/* 149 * Arithmetic functions. 150 * Each of these may modify its inputs (f1,f2) and/or the temporary. 151 * Each returns a pointer to the result and/or sets exceptions. 152 */ 153struct fpn *fpu_add(struct fpemu *); 154#define fpu_sub(fe) ((fe)->fe_f2.fp_sign ^= 1, fpu_add(fe)) 155struct fpn *fpu_mul(struct fpemu *); 156struct fpn *fpu_div(struct fpemu *); 157struct fpn *fpu_sqrt(struct fpemu *); 158 159/* 160 * Other functions. 161 */ 162 163/* Perform a compare instruction (with or without unordered exception). */ 164void fpu_compare(struct fpemu *, int); 165 166/* Build a new Quiet NaN (sign=0, frac=all 1's). */ 167struct fpn *fpu_newnan(struct fpemu *); 168 169/* 170 * Shift a number right some number of bits, taking care of round/sticky. 171 * Note that the result is probably not a well-formed number (it will lack 172 * the normal 1-bit mant[0]&FP_1). 173 */ 174int fpu_shr(struct fpn *, int); 175 176void fpu_explode(struct fpemu *, struct fpn *, int, int); 177void fpu_implode(struct fpemu *, struct fpn *, int, u_int *); 178 179#ifdef DEBUG 180#define FPE_EX 0x1 181#define FPE_INSN 0x2 182#define FPE_OP 0x4 183#define FPE_REG 0x8 184extern int fpe_debug; 185void fpu_dumpfpn(struct fpn *); 186#define DPRINTF(x, y) if (fpe_debug & (x)) printf y 187#define DUMPFPN(x, f) if (fpe_debug & (x)) fpu_dumpfpn((f)) 188#else 189#define DPRINTF(x, y) 190#define DUMPFPN(x, f) 191#endif 192