xref: /netbsd/sys/arch/sparc/fpu/fpu_explode.c (revision 93562b35)
1 /*	$NetBSD: fpu_explode.c,v 1.13 2022/08/28 22:09:26 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_explode.c	8.1 (Berkeley) 6/11/93
41  */
42 
43 /*
44  * FPU subroutines: `explode' the machine's `packed binary' format numbers
45  * into our internal format.
46  */
47 
48 #include <sys/cdefs.h>
49 __KERNEL_RCSID(0, "$NetBSD: fpu_explode.c,v 1.13 2022/08/28 22:09:26 rin Exp $");
50 
51 #if defined(_KERNEL_OPT)
52 #include "opt_sparc_arch.h"
53 #endif
54 
55 #include <sys/types.h>
56 #include <sys/systm.h>
57 
58 #include <machine/ieee.h>
59 #include <machine/instr.h>
60 #include <machine/reg.h>
61 
62 #include <sparc/fpu/fpu_arith.h>
63 #include <sparc/fpu/fpu_emu.h>
64 #include <sparc/fpu/fpu_extern.h>
65 
66 /*
67  * N.B.: in all of the following, we assume the FP format is
68  *
69  *	---------------------------
70  *	| s | exponent | fraction |
71  *	---------------------------
72  *
73  * (which represents -1**s * 1.fraction * 2**exponent), so that the
74  * sign bit is way at the top (bit 31), the exponent is next, and
75  * then the remaining bits mark the fraction.  A zero exponent means
76  * zero or denormalized (0.fraction rather than 1.fraction), and the
77  * maximum possible exponent, 2bias+1, signals inf (fraction==0) or NaN.
78  *
79  * Since the sign bit is always the topmost bit---this holds even for
80  * integers---we set that outside all the *tof functions.  Each function
81  * returns the class code for the new number (but note that we use
82  * FPC_QNAN for all NaNs; fpu_explode will fix this if appropriate).
83  */
84 
85 /*
86  * int -> fpn.
87  */
88 int
fpu_itof(struct fpn * fp,u_int i)89 fpu_itof(struct fpn *fp, u_int i)
90 {
91 
92 	if (i == 0)
93 		return (FPC_ZERO);
94 	/*
95 	 * The value FP_1 represents 2^FP_LG, so set the exponent
96 	 * there and let normalization fix it up.  Convert negative
97 	 * numbers to sign-and-magnitude.  Note that this relies on
98 	 * fpu_norm()'s handling of `supernormals'; see fpu_subr.c.
99 	 */
100 	fp->fp_exp = FP_LG;
101 	fp->fp_mant[0] = (int)i < 0 ? -i : i;
102 	fp->fp_mant[1] = 0;
103 	fp->fp_mant[2] = 0;
104 	fp->fp_mant[3] = 0;
105 	fpu_norm(fp);
106 	return (FPC_NUM);
107 }
108 
109 #ifdef SUN4U
110 /*
111  * 64-bit int -> fpn.
112  */
113 int
fpu_xtof(struct fpn * fp,uint64_t i)114 fpu_xtof(struct fpn *fp, uint64_t i)
115 {
116 
117 	if (i == 0)
118 		return (FPC_ZERO);
119 	/*
120 	 * The value FP_1 represents 2^FP_LG, so set the exponent
121 	 * there and let normalization fix it up.  Convert negative
122 	 * numbers to sign-and-magnitude.  Note that this relies on
123 	 * fpu_norm()'s handling of `supernormals'; see fpu_subr.c.
124 	 */
125 	fp->fp_exp = FP_LG2;
126 	*((int64_t*)fp->fp_mant) = (int64_t)i < 0 ? -i : i;
127 	fp->fp_mant[2] = 0;
128 	fp->fp_mant[3] = 0;
129 	fpu_norm(fp);
130 	return (FPC_NUM);
131 }
132 #endif /* SUN4U */
133 
134 #define	mask(nbits) ((1L << (nbits)) - 1)
135 
136 /*
137  * All external floating formats convert to internal in the same manner,
138  * as defined here.  Note that only normals get an implied 1.0 inserted.
139  */
140 #define	FP_TOF(exp, expbias, allfrac, f0, f1, f2, f3) \
141 	if (exp == 0) { \
142 		if (allfrac == 0) \
143 			return (FPC_ZERO); \
144 		fp->fp_exp = 1 - expbias; \
145 		fp->fp_mant[0] = f0; \
146 		fp->fp_mant[1] = f1; \
147 		fp->fp_mant[2] = f2; \
148 		fp->fp_mant[3] = f3; \
149 		fpu_norm(fp); \
150 		return (FPC_NUM); \
151 	} \
152 	if (exp == (2 * expbias + 1)) { \
153 		if (allfrac == 0) \
154 			return (FPC_INF); \
155 		fp->fp_mant[0] = f0; \
156 		fp->fp_mant[1] = f1; \
157 		fp->fp_mant[2] = f2; \
158 		fp->fp_mant[3] = f3; \
159 		return (FPC_QNAN); \
160 	} \
161 	fp->fp_exp = exp - expbias; \
162 	fp->fp_mant[0] = FP_1 | f0; \
163 	fp->fp_mant[1] = f1; \
164 	fp->fp_mant[2] = f2; \
165 	fp->fp_mant[3] = f3; \
166 	return (FPC_NUM)
167 
168 /*
169  * 32-bit single precision -> fpn.
170  * We assume a single occupies at most (64-FP_LG) bits in the internal
171  * format: i.e., needs at most fp_mant[0] and fp_mant[1].
172  */
173 int
fpu_stof(struct fpn * fp,u_int i)174 fpu_stof(struct fpn *fp, u_int i)
175 {
176 	int exp;
177 	u_int frac, f0, f1;
178 #define SNG_SHIFT (SNG_FRACBITS - FP_LG)
179 
180 	exp = (i >> (32 - 1 - SNG_EXPBITS)) & mask(SNG_EXPBITS);
181 	frac = i & mask(SNG_FRACBITS);
182 	f0 = frac >> SNG_SHIFT;
183 	f1 = frac << (32 - SNG_SHIFT);
184 	FP_TOF(exp, SNG_EXP_BIAS, frac, f0, f1, 0, 0);
185 }
186 
187 /*
188  * 64-bit double -> fpn.
189  * We assume this uses at most (96-FP_LG) bits.
190  */
191 int
fpu_dtof(struct fpn * fp,u_int i,u_int j)192 fpu_dtof(struct fpn *fp, u_int i, u_int j)
193 {
194 	int exp;
195 	u_int frac, f0, f1, f2;
196 #define DBL_SHIFT (DBL_FRACBITS - 32 - FP_LG)
197 
198 	exp = (i >> (32 - 1 - DBL_EXPBITS)) & mask(DBL_EXPBITS);
199 	frac = i & mask(DBL_FRACBITS - 32);
200 	f0 = frac >> DBL_SHIFT;
201 	f1 = (frac << (32 - DBL_SHIFT)) | (j >> DBL_SHIFT);
202 	f2 = j << (32 - DBL_SHIFT);
203 	frac |= j;
204 	FP_TOF(exp, DBL_EXP_BIAS, frac, f0, f1, f2, 0);
205 }
206 
207 /*
208  * 128-bit extended -> fpn.
209  */
210 int
fpu_qtof(struct fpn * fp,u_int i,u_int j,u_int k,u_int l)211 fpu_qtof(struct fpn *fp, u_int i, u_int j, u_int k, u_int l)
212 {
213 	int exp;
214 	u_int frac, f0, f1, f2, f3;
215 #define EXT_SHIFT (-(EXT_FRACBITS - 3 * 32 - FP_LG))	/* left shift! */
216 
217 	/*
218 	 * Note that ext and fpn `line up', hence no shifting needed.
219 	 */
220 	exp = (i >> (32 - 1 - EXT_EXPBITS)) & mask(EXT_EXPBITS);
221 	frac = i & mask(EXT_FRACBITS - 3 * 32);
222 	f0 = (frac << EXT_SHIFT) | (j >> (32 - EXT_SHIFT));
223 	f1 = (j << EXT_SHIFT) | (k >> (32 - EXT_SHIFT));
224 	f2 = (k << EXT_SHIFT) | (l >> (32 - EXT_SHIFT));
225 	f3 = l << EXT_SHIFT;
226 	frac |= j | k | l;
227 	FP_TOF(exp, EXT_EXP_BIAS, frac, f0, f1, f2, f3);
228 }
229 
230 /*
231  * Explode the contents of a register / regpair / regquad.
232  * If the input is a signalling NaN, an NV (invalid) exception
233  * will be set.  (Note that nothing but NV can occur until ALU
234  * operations are performed.)
235  */
236 void
fpu_explode(struct fpemu * fe,struct fpn * fp,int type,int reg)237 fpu_explode(struct fpemu *fe, struct fpn *fp, int type, int reg)
238 {
239 	u_int s, *space;
240 #ifdef SUN4U
241 	uint64_t l, *xspace;
242 
243 	xspace = (uint64_t *)&fe->fe_fpstate->fs_regs[reg & ~1];
244 	l = xspace[0];
245 #endif /* SUN4U */
246 	space = &fe->fe_fpstate->fs_regs[reg];
247 	s = space[0];
248 	fp->fp_sign = s >> 31;
249 	fp->fp_sticky = 0;
250 	switch (type) {
251 #ifdef SUN4U
252 	case FTYPE_LNG:
253 		s = fpu_xtof(fp, l);
254 		break;
255 #endif /* SUN4U */
256 
257 	case FTYPE_INT:
258 		s = fpu_itof(fp, s);
259 		break;
260 
261 	case FTYPE_SNG:
262 		s = fpu_stof(fp, s);
263 		break;
264 
265 	case FTYPE_DBL:
266 		s = fpu_dtof(fp, s, space[1]);
267 		break;
268 
269 	case FTYPE_EXT:
270 		s = fpu_qtof(fp, s, space[1], space[2], space[3]);
271 		break;
272 
273 	default:
274 		panic("fpu_explode");
275 	}
276 
277 	if (s == FPC_QNAN && (fp->fp_mant[0] & FP_QUIETBIT) == 0) {
278 		/*
279 		 * Input is a signalling NaN.  All operations that return
280 		 * an input NaN operand put it through a ``NaN conversion'',
281 		 * which basically just means ``turn on the quiet bit''.
282 		 * We do this here so that all NaNs internally look quiet
283 		 * (we can tell signalling ones by their class).
284 		 */
285 		fp->fp_mant[0] |= FP_QUIETBIT;
286 		fe->fe_cx = FSR_NV;	/* assert invalid operand */
287 		s = FPC_SNAN;
288 	}
289 	fp->fp_class = s;
290 	DPRINTF(FPE_REG, ("fpu_explode: %%%c%d => ", (type == FTYPE_LNG) ? 'x' :
291 		((type == FTYPE_INT) ? 'i' :
292 			((type == FTYPE_SNG) ? 's' :
293 				((type == FTYPE_DBL) ? 'd' :
294 					((type == FTYPE_EXT) ? 'q' : '?')))),
295 		reg));
296 #ifdef DEBUG
297 	if (fpe_debug & FPE_REG) {
298 		if (type == FTYPE_INT) printf("%d ", s);
299 #ifdef SUN4U
300 #ifdef _LP64
301 		if (type == FTYPE_LNG) printf("%ld ", l);
302 #else
303 		if (type == FTYPE_LNG) printf("%lld ", l);
304 #endif
305 #endif /* SUN4U */
306 	}
307 #endif /* DEBUG */
308 	DUMPFPN(FPE_REG, fp);
309 	DPRINTF(FPE_REG, ("\n"));
310 }
311