1 /*        $NetBSD: fpu_explode.c,v 1.14 2022/09/07 06:51:58 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 1.14 2022/09/07 06:51:58 rin Exp $");
50 
51 #include <sys/types.h>
52 #include <sys/systm.h>
53 
54 #include <powerpc/instr.h>
55 #include <machine/fpu.h>
56 #include <machine/ieee.h>
57 #include <machine/reg.h>
58 
59 #include <powerpc/fpu/fpu_arith.h>
60 #include <powerpc/fpu/fpu_emu.h>
61 #include <powerpc/fpu/fpu_extern.h>
62 
63 static int fpu_itof(struct fpn *, u_int);
64 static int fpu_xtof(struct fpn *, uint64_t);
65 static int fpu_stof(struct fpn *, u_int);
66 static int fpu_dtof(struct fpn *, u_int, u_int);
67 
68 /*
69  * N.B.: in all of the following, we assume the FP format is
70  *
71  *        ---------------------------
72  *        | s | exponent | fraction |
73  *        ---------------------------
74  *
75  * (which represents -1**s * 1.fraction * 2**exponent), so that the
76  * sign bit is way at the top (bit 31), the exponent is next, and
77  * then the remaining bits mark the fraction.  A zero exponent means
78  * zero or denormalized (0.fraction rather than 1.fraction), and the
79  * maximum possible exponent, 2bias+1, signals inf (fraction==0) or NaN.
80  *
81  * Since the sign bit is always the topmost bit---this holds even for
82  * integers---we set that outside all the *tof functions.  Each function
83  * returns the class code for the new number (but note that we use
84  * FPC_QNAN for all NaNs; fpu_explode will fix this if appropriate).
85  */
86 
87 /*
88  * int -> fpn.
89  */
90 static int
fpu_itof(struct fpn * fp,u_int lo)91 fpu_itof(struct fpn *fp, u_int lo)
92 {
93 
94           if (lo == 0)
95                     return (FPC_ZERO);
96           /*
97            * The value FP_1 represents 2^FP_LG, so set the exponent
98            * there and let normalization fix it up.  Convert negative
99            * numbers to sign-and-magnitude.  Note that this relies on
100            * fpu_norm()'s handling of `supernormals'; see fpu_subr.c.
101            */
102           fp->fp_exp = FP_LG;
103           fp->fp_mant[0] = (int)lo < 0 ? -lo : lo;
104           fp->fp_mant[1] = 0;
105           fp->fp_mant[2] = 0;
106           fp->fp_mant[3] = 0;
107           fpu_norm(fp);
108           return (FPC_NUM);
109 }
110 
111 /*
112  * 64-bit int -> fpn.
113  */
114 static int
fpu_xtof(struct fpn * fp,uint64_t i)115 fpu_xtof(struct fpn *fp, uint64_t i)
116 {
117 
118           if (i == 0)
119                     return (FPC_ZERO);
120           /*
121            * The value FP_1 represents 2^FP_LG, so set the exponent
122            * there and let normalization fix it up.  Convert negative
123            * numbers to sign-and-magnitude.  Note that this relies on
124            * fpu_norm()'s handling of `supernormals'; see fpu_subr.c.
125            */
126           fp->fp_exp = FP_LG2;
127           *((int64_t*)fp->fp_mant) = (int64_t)i < 0 ? -i : i;
128           fp->fp_mant[2] = 0;
129           fp->fp_mant[3] = 0;
130           fpu_norm(fp);
131           return (FPC_NUM);
132 }
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 static int
fpu_stof(struct fpn * fp,u_int hi)174 fpu_stof(struct fpn *fp, u_int hi)
175 {
176           int exp;
177           u_int frac, f0, f1;
178 #define SNG_SHIFT (SNG_FRACBITS - FP_LG)
179 
180           exp = (hi >> (32 - 1 - SNG_EXPBITS)) & mask(SNG_EXPBITS);
181           frac = hi & 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 static int
fpu_dtof(struct fpn * fp,u_int hi,u_int lo)192 fpu_dtof(struct fpn *fp, u_int hi, u_int lo)
193 {
194           int exp;
195           u_int frac, f0, f1, f2;
196 #define DBL_SHIFT (DBL_FRACBITS - 32 - FP_LG)
197 
198           exp = (hi >> (32 - 1 - DBL_EXPBITS)) & mask(DBL_EXPBITS);
199           frac = hi & mask(DBL_FRACBITS - 32);
200           f0 = frac >> DBL_SHIFT;
201           f1 = (frac << (32 - DBL_SHIFT)) | (lo >> DBL_SHIFT);
202           f2 = lo << (32 - DBL_SHIFT);
203           frac |= lo;
204           FP_TOF(exp, DBL_EXP_BIAS, frac, f0, f1, f2, 0);
205 }
206 
207 /*
208  * Explode the contents of a register / regpair / regquad.
209  * If the input is a signalling NaN, an NV (invalid) exception
210  * will be set.  (Note that nothing but NV can occur until ALU
211  * operations are performed.)
212  */
213 void
fpu_explode(struct fpemu * fe,struct fpn * fp,int type,uint64_t i)214 fpu_explode(struct fpemu *fe, struct fpn *fp, int type, uint64_t i)
215 {
216           u_int hi, lo;
217           int class;
218 
219           hi = (u_int)(i >> 32);
220           lo = (u_int)i;
221           fp->fp_sign = hi >> 31;
222           fp->fp_sticky = 0;
223           switch (type) {
224 
225           case FTYPE_LNG:
226                     class = fpu_xtof(fp, i);
227                     break;
228 
229           case FTYPE_INT:
230                     fp->fp_sign = lo >> 31;
231                     class = fpu_itof(fp, lo);
232                     break;
233 
234           case FTYPE_SNG:
235                     class = fpu_stof(fp, hi);
236                     break;
237 
238           case FTYPE_DBL:
239                     class = fpu_dtof(fp, hi, lo);
240                     break;
241 
242           default:
243                     panic("fpu_explode: invalid type %d", type);
244           }
245 
246           if (class == FPC_QNAN && (fp->fp_mant[0] & FP_QUIETBIT) == 0) {
247                     /*
248                      * Input is a signalling NaN.  All operations that return
249                      * an input NaN operand put it through a ``NaN conversion'',
250                      * which basically just means ``turn on the quiet bit''.
251                      * We do this here so that all NaNs internally look quiet
252                      * (we can tell signalling ones by their class).
253                      */
254                     fp->fp_mant[0] |= FP_QUIETBIT;
255                     fe->fe_cx = FPSCR_VXSNAN;     /* assert invalid operand */
256                     class = FPC_SNAN;
257           }
258           fp->fp_class = class;
259           DUMPFPN(FPE_REG, fp);
260 }
261