0
|
1 /* Signed and unsigned multiplication and division and modulus for CRIS.
|
|
2 Contributed by Axis Communications.
|
|
3 Written by Hans-Peter Nilsson <hp@axis.se>, c:a 1992.
|
|
4
|
|
5 Copyright (C) 1998, 1999, 2000, 2001, 2002,
|
|
6 2005, 2009 Free Software Foundation, Inc.
|
|
7
|
|
8 This file is part of GCC.
|
|
9
|
|
10 GCC is free software; you can redistribute it and/or modify it
|
|
11 under the terms of the GNU General Public License as published by the
|
|
12 Free Software Foundation; either version 3, or (at your option) any
|
|
13 later version.
|
|
14
|
|
15 This file is distributed in the hope that it will be useful, but
|
|
16 WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
18 General Public License for more details.
|
|
19
|
|
20 Under Section 7 of GPL version 3, you are granted additional
|
|
21 permissions described in the GCC Runtime Library Exception, version
|
|
22 3.1, as published by the Free Software Foundation.
|
|
23
|
|
24 You should have received a copy of the GNU General Public License and
|
|
25 a copy of the GCC Runtime Library Exception along with this program;
|
|
26 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
27 <http://www.gnu.org/licenses/>. */
|
|
28
|
|
29
|
|
30 /* Note that we provide prototypes for all "const" functions, to attach
|
|
31 the const attribute. This is necessary in 2.7.2 - adding the
|
|
32 attribute to the function *definition* is a syntax error.
|
|
33 This did not work with e.g. 2.1; back then, the return type had to
|
|
34 be "const". */
|
|
35
|
|
36 #include "config.h"
|
|
37
|
|
38 #if defined (__CRIS_arch_version) && __CRIS_arch_version >= 3
|
|
39 #define LZ(v) __builtin_clz (v)
|
|
40 #endif
|
|
41
|
|
42
|
|
43 #if defined (L_udivsi3) || defined (L_divsi3) || defined (L_umodsi3) \
|
|
44 || defined (L_modsi3)
|
|
45 /* Result type of divmod worker function. */
|
|
46 struct quot_rem
|
|
47 {
|
|
48 long quot;
|
|
49 long rem;
|
|
50 };
|
|
51
|
|
52 /* This is the worker function for div and mod. It is inlined into the
|
|
53 respective library function. Parameter A must have bit 31 == 0. */
|
|
54
|
|
55 static __inline__ struct quot_rem
|
|
56 do_31div (unsigned long a, unsigned long b)
|
|
57 __attribute__ ((__const__, __always_inline__));
|
|
58
|
|
59 static __inline__ struct quot_rem
|
|
60 do_31div (unsigned long a, unsigned long b)
|
|
61 {
|
|
62 /* Adjust operands and result if a is 31 bits. */
|
|
63 long extra = 0;
|
|
64 int quot_digits = 0;
|
|
65
|
|
66 if (b == 0)
|
|
67 {
|
|
68 struct quot_rem ret;
|
|
69 ret.quot = 0xffffffff;
|
|
70 ret.rem = 0xffffffff;
|
|
71 return ret;
|
|
72 }
|
|
73
|
|
74 if (a < b)
|
|
75 return (struct quot_rem) { 0, a };
|
|
76
|
|
77 #ifdef LZ
|
|
78 if (b <= a)
|
|
79 {
|
|
80 quot_digits = LZ (b) - LZ (a);
|
|
81 quot_digits += (a >= (b << quot_digits));
|
|
82 b <<= quot_digits;
|
|
83 }
|
|
84 #else
|
|
85 while (b <= a)
|
|
86 {
|
|
87 b <<= 1;
|
|
88 quot_digits++;
|
|
89 }
|
|
90 #endif
|
|
91
|
|
92 /* Is a 31 bits? Note that bit 31 is handled by the caller. */
|
|
93 if (a & 0x40000000)
|
|
94 {
|
|
95 /* Then make b:s highest bit max 0x40000000, because it must have
|
|
96 been 0x80000000 to be 1 bit higher than a. */
|
|
97 b >>= 1;
|
|
98
|
|
99 /* Adjust a to be maximum 0x3fffffff, i.e. two upper bits zero. */
|
|
100 if (a >= b)
|
|
101 {
|
|
102 a -= b;
|
|
103 extra = 1 << (quot_digits - 1);
|
|
104 }
|
|
105 else
|
|
106 {
|
|
107 a -= b >> 1;
|
|
108
|
|
109 /* Remember that we adjusted a by subtracting b * 2 ** Something. */
|
|
110 extra = 1 << quot_digits;
|
|
111 }
|
|
112
|
|
113 /* The number of quotient digits will be one less, because
|
|
114 we just adjusted b. */
|
|
115 quot_digits--;
|
|
116 }
|
|
117
|
|
118 /* Now do the division part. */
|
|
119
|
|
120 /* Subtract b and add ones to the right when a >= b
|
|
121 i.e. "a - (b - 1) == (a - b) + 1". */
|
|
122 b--;
|
|
123
|
|
124 #define DS __asm__ ("dstep %2,%0" : "=r" (a) : "0" (a), "r" (b))
|
|
125
|
|
126 switch (quot_digits)
|
|
127 {
|
|
128 case 32: DS; case 31: DS; case 30: DS; case 29: DS;
|
|
129 case 28: DS; case 27: DS; case 26: DS; case 25: DS;
|
|
130 case 24: DS; case 23: DS; case 22: DS; case 21: DS;
|
|
131 case 20: DS; case 19: DS; case 18: DS; case 17: DS;
|
|
132 case 16: DS; case 15: DS; case 14: DS; case 13: DS;
|
|
133 case 12: DS; case 11: DS; case 10: DS; case 9: DS;
|
|
134 case 8: DS; case 7: DS; case 6: DS; case 5: DS;
|
|
135 case 4: DS; case 3: DS; case 2: DS; case 1: DS;
|
|
136 case 0:;
|
|
137 }
|
|
138
|
|
139 {
|
|
140 struct quot_rem ret;
|
|
141 ret.quot = (a & ((1 << quot_digits) - 1)) + extra;
|
|
142 ret.rem = a >> quot_digits;
|
|
143 return ret;
|
|
144 }
|
|
145 }
|
|
146
|
|
147 #ifdef L_udivsi3
|
|
148 unsigned long
|
|
149 __Udiv (unsigned long a, unsigned long b) __attribute__ ((__const__));
|
|
150
|
|
151 unsigned long
|
|
152 __Udiv (unsigned long a, unsigned long b)
|
|
153 {
|
|
154 long extra = 0;
|
|
155
|
|
156 /* Adjust operands and result, if a and/or b is 32 bits. */
|
|
157 /* Effectively: b & 0x80000000. */
|
|
158 if ((long) b < 0)
|
|
159 return a >= b;
|
|
160
|
|
161 /* Effectively: a & 0x80000000. */
|
|
162 if ((long) a < 0)
|
|
163 {
|
|
164 int tmp = 0;
|
|
165
|
|
166 if (b == 0)
|
|
167 return 0xffffffff;
|
|
168 #ifdef LZ
|
|
169 tmp = LZ (b);
|
|
170 #else
|
|
171 for (tmp = 31; (((long) b & (1 << tmp)) == 0); tmp--)
|
|
172 ;
|
|
173
|
|
174 tmp = 31 - tmp;
|
|
175 #endif
|
|
176
|
|
177 if ((b << tmp) > a)
|
|
178 {
|
|
179 extra = 1 << (tmp-1);
|
|
180 a -= b << (tmp - 1);
|
|
181 }
|
|
182 else
|
|
183 {
|
|
184 extra = 1 << tmp;
|
|
185 a -= b << tmp;
|
|
186 }
|
|
187 }
|
|
188
|
|
189 return do_31div (a, b).quot+extra;
|
|
190 }
|
|
191 #endif /* L_udivsi3 */
|
|
192
|
|
193 #ifdef L_divsi3
|
|
194 long
|
|
195 __Div (long a, long b) __attribute__ ((__const__));
|
|
196
|
|
197 long
|
|
198 __Div (long a, long b)
|
|
199 {
|
|
200 long extra = 0;
|
|
201 long sign = (b < 0) ? -1 : 1;
|
|
202
|
|
203 /* We need to handle a == -2147483648 as expected and must while
|
|
204 doing that avoid producing a sequence like "abs (a) < 0" as GCC
|
|
205 may optimize out the test. That sequence may not be obvious as
|
|
206 we call inline functions. Testing for a being negative and
|
|
207 handling (presumably much rarer than positive) enables us to get
|
|
208 a bit of optimization for an (accumulated) reduction of the
|
|
209 penalty of the 0x80000000 special-case. */
|
|
210 if (a < 0)
|
|
211 {
|
|
212 sign = -sign;
|
|
213
|
|
214 if ((a & 0x7fffffff) == 0)
|
|
215 {
|
|
216 /* We're at 0x80000000. Tread carefully. */
|
|
217 a -= b * sign;
|
|
218 extra = sign;
|
|
219 }
|
|
220 a = -a;
|
|
221 }
|
|
222
|
|
223 /* We knowingly penalize pre-v10 models by multiplication with the
|
|
224 sign. */
|
|
225 return sign * do_31div (a, __builtin_labs (b)).quot + extra;
|
|
226 }
|
|
227 #endif /* L_divsi3 */
|
|
228
|
|
229
|
|
230 #ifdef L_umodsi3
|
|
231 unsigned long
|
|
232 __Umod (unsigned long a, unsigned long b) __attribute__ ((__const__));
|
|
233
|
|
234 unsigned long
|
|
235 __Umod (unsigned long a, unsigned long b)
|
|
236 {
|
|
237 /* Adjust operands and result if a and/or b is 32 bits. */
|
|
238 if ((long) b < 0)
|
|
239 return a >= b ? a - b : a;
|
|
240
|
|
241 if ((long) a < 0)
|
|
242 {
|
|
243 int tmp = 0;
|
|
244
|
|
245 if (b == 0)
|
|
246 return a;
|
|
247 #ifdef LZ
|
|
248 tmp = LZ (b);
|
|
249 #else
|
|
250 for (tmp = 31; (((long) b & (1 << tmp)) == 0); tmp--)
|
|
251 ;
|
|
252 tmp = 31 - tmp;
|
|
253 #endif
|
|
254
|
|
255 if ((b << tmp) > a)
|
|
256 {
|
|
257 a -= b << (tmp - 1);
|
|
258 }
|
|
259 else
|
|
260 {
|
|
261 a -= b << tmp;
|
|
262 }
|
|
263 }
|
|
264
|
|
265 return do_31div (a, b).rem;
|
|
266 }
|
|
267 #endif /* L_umodsi3 */
|
|
268
|
|
269 #ifdef L_modsi3
|
|
270 long
|
|
271 __Mod (long a, long b) __attribute__ ((__const__));
|
|
272
|
|
273 long
|
|
274 __Mod (long a, long b)
|
|
275 {
|
|
276 long sign = 1;
|
|
277
|
|
278 /* We need to handle a == -2147483648 as expected and must while
|
|
279 doing that avoid producing a sequence like "abs (a) < 0" as GCC
|
|
280 may optimize out the test. That sequence may not be obvious as
|
|
281 we call inline functions. Testing for a being negative and
|
|
282 handling (presumably much rarer than positive) enables us to get
|
|
283 a bit of optimization for an (accumulated) reduction of the
|
|
284 penalty of the 0x80000000 special-case. */
|
|
285 if (a < 0)
|
|
286 {
|
|
287 sign = -1;
|
|
288 if ((a & 0x7fffffff) == 0)
|
|
289 /* We're at 0x80000000. Tread carefully. */
|
|
290 a += __builtin_labs (b);
|
|
291 a = -a;
|
|
292 }
|
|
293
|
|
294 return sign * do_31div (a, __builtin_labs (b)).rem;
|
|
295 }
|
|
296 #endif /* L_modsi3 */
|
|
297 #endif /* L_udivsi3 || L_divsi3 || L_umodsi3 || L_modsi3 */
|
|
298
|
|
299 /*
|
|
300 * Local variables:
|
|
301 * eval: (c-set-style "gnu")
|
|
302 * indent-tabs-mode: t
|
|
303 * End:
|
|
304 */
|