| 1 |
/* Signed and unsigned multiplication and division and modulus for CRIS. |
| 2 |
Contributed by Axis Communications. |
| 3 |
Written by Hans-Peter Nilsson <[email protected]>, c:a 1992. |
| 4 |
|
| 5 |
Copyright (C) 1998-2025 Free Software Foundation, Inc. |
| 6 |
|
| 7 |
This file is part of GCC. |
| 8 |
|
| 9 |
GCC is free software; you can redistribute it and/or modify it |
| 10 |
under the terms of the GNU General Public License as published by the |
| 11 |
Free Software Foundation; either version 3, or (at your option) any |
| 12 |
later version. |
| 13 |
|
| 14 |
This file is distributed in the hope that it will be useful, but |
| 15 |
WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 17 |
General Public License for more details. |
| 18 |
|
| 19 |
Under Section 7 of GPL version 3, you are granted additional |
| 20 |
permissions described in the GCC Runtime Library Exception, version |
| 21 |
3.1, as published by the Free Software Foundation. |
| 22 |
|
| 23 |
You should have received a copy of the GNU General Public License and |
| 24 |
a copy of the GCC Runtime Library Exception along with this program; |
| 25 |
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
| 26 |
<http://www.gnu.org/licenses/>. */ |
| 27 |
|
| 28 |
|
| 29 |
/* Note that we provide prototypes for all "const" functions, to attach |
| 30 |
the const attribute. This is necessary in 2.7.2 - adding the |
| 31 |
attribute to the function *definition* is a syntax error. |
| 32 |
This did not work with e.g. 2.1; back then, the return type had to |
| 33 |
be "const". */ |
| 34 |
|
| 35 |
#include "config.h" |
| 36 |
|
| 37 |
#if defined (__CRIS_arch_version) && __CRIS_arch_version >= 3 |
| 38 |
#define LZ(v) __builtin_clz (v) |
| 39 |
#endif |
| 40 |
|
| 41 |
/* In (at least) the 4.7 series, GCC doesn't automatically choose the |
| 42 |
most optimal strategy, possibly related to insufficient modelling of |
| 43 |
delay-slot costs. */ |
| 44 |
#if defined (__CRIS_arch_version) && __CRIS_arch_version >= 10 |
| 45 |
#define SIGNMULT(s, a) ((s) * (a)) /* Cheap multiplication, better than branch. */ |
| 46 |
#else |
| 47 |
#define SIGNMULT(s, a) ((s) < 0 ? -(a) : (a)) /* Branches are still better. */ |
| 48 |
#endif |
| 49 |
|
| 50 |
#if defined (L_udivsi3) || defined (L_divsi3) || defined (L_umodsi3) \ |
| 51 |
|| defined (L_modsi3) |
| 52 |
/* Result type of divmod worker function. */ |
| 53 |
struct quot_rem |
| 54 |
{ |
| 55 |
long quot; |
| 56 |
long rem; |
| 57 |
}; |
| 58 |
|
| 59 |
/* This is the worker function for div and mod. It is inlined into the |
| 60 |
respective library function. Parameter A must have bit 31 == 0. */ |
| 61 |
|
| 62 |
static __inline__ struct quot_rem |
| 63 |
do_31div (unsigned long a, unsigned long b) |
| 64 |
__attribute__ ((__const__, __always_inline__)); |
| 65 |
|
| 66 |
static __inline__ struct quot_rem |
| 67 |
do_31div (unsigned long a, unsigned long b) |
| 68 |
{ |
| 69 |
/* Adjust operands and result if a is 31 bits. */ |
| 70 |
long extra = 0; |
| 71 |
int quot_digits = 0; |
| 72 |
|
| 73 |
if (b == 0) |
| 74 |
{ |
| 75 |
struct quot_rem ret; |
| 76 |
ret.quot = 0xffffffff; |
| 77 |
ret.rem = 0xffffffff; |
| 78 |
return ret; |
| 79 |
} |
| 80 |
|
| 81 |
if (a < b) |
| 82 |
return (struct quot_rem) { 0, a }; |
| 83 |
|
| 84 |
#ifdef LZ |
| 85 |
if (b <= a) |
| 86 |
{ |
| 87 |
quot_digits = LZ (b) - LZ (a); |
| 88 |
quot_digits += (a >= (b << quot_digits)); |
| 89 |
b <<= quot_digits; |
| 90 |
} |
| 91 |
#else |
| 92 |
while (b <= a) |
| 93 |
{ |
| 94 |
b <<= 1; |
| 95 |
quot_digits++; |
| 96 |
} |
| 97 |
#endif |
| 98 |
|
| 99 |
/* Is a 31 bits? Note that bit 31 is handled by the caller. */ |
| 100 |
if (a & 0x40000000) |
| 101 |
{ |
| 102 |
/* Then make b:s highest bit max 0x40000000, because it must have |
| 103 |
been 0x80000000 to be 1 bit higher than a. */ |
| 104 |
b >>= 1; |
| 105 |
|
| 106 |
/* Adjust a to be maximum 0x3fffffff, i.e. two upper bits zero. */ |
| 107 |
if (a >= b) |
| 108 |
{ |
| 109 |
a -= b; |
| 110 |
extra = 1 << (quot_digits - 1); |
| 111 |
} |
| 112 |
else |
| 113 |
{ |
| 114 |
a -= b >> 1; |
| 115 |
|
| 116 |
/* Remember that we adjusted a by subtracting b * 2 ** Something. */ |
| 117 |
extra = 1 << quot_digits; |
| 118 |
} |
| 119 |
|
| 120 |
/* The number of quotient digits will be one less, because |
| 121 |
we just adjusted b. */ |
| 122 |
quot_digits--; |
| 123 |
} |
| 124 |
|
| 125 |
/* Now do the division part. */ |
| 126 |
|
| 127 |
/* Subtract b and add ones to the right when a >= b |
| 128 |
i.e. "a - (b - 1) == (a - b) + 1". */ |
| 129 |
b--; |
| 130 |
|
| 131 |
#define DS __asm__ ("dstep %2,%0" : "=r" (a) : "0" (a), "r" (b)); \ |
| 132 |
__attribute__ ((__fallthrough__)) |
| 133 |
|
| 134 |
switch (quot_digits) |
| 135 |
{ |
| 136 |
case 32: DS; case 31: DS; case 30: DS; case 29: DS; |
| 137 |
case 28: DS; case 27: DS; case 26: DS; case 25: DS; |
| 138 |
case 24: DS; case 23: DS; case 22: DS; case 21: DS; |
| 139 |
case 20: DS; case 19: DS; case 18: DS; case 17: DS; |
| 140 |
case 16: DS; case 15: DS; case 14: DS; case 13: DS; |
| 141 |
case 12: DS; case 11: DS; case 10: DS; case 9: DS; |
| 142 |
case 8: DS; case 7: DS; case 6: DS; case 5: DS; |
| 143 |
case 4: DS; case 3: DS; case 2: DS; case 1: DS; |
| 144 |
case 0:; |
| 145 |
} |
| 146 |
|
| 147 |
{ |
| 148 |
struct quot_rem ret; |
| 149 |
ret.quot = (a & ((1 << quot_digits) - 1)) + extra; |
| 150 |
ret.rem = a >> quot_digits; |
| 151 |
return ret; |
| 152 |
} |
| 153 |
} |
| 154 |
|
| 155 |
#ifdef L_udivsi3 |
| 156 |
unsigned long |
| 157 |
__Udiv (unsigned long a, unsigned long b) __attribute__ ((__const__)); |
| 158 |
|
| 159 |
unsigned long |
| 160 |
__Udiv (unsigned long a, unsigned long b) |
| 161 |
{ |
| 162 |
long extra = 0; |
| 163 |
|
| 164 |
/* Adjust operands and result, if a and/or b is 32 bits. */ |
| 165 |
/* Effectively: b & 0x80000000. */ |
| 166 |
if ((long) b < 0) |
| 167 |
return a >= b; |
| 168 |
|
| 169 |
/* Effectively: a & 0x80000000. */ |
| 170 |
if ((long) a < 0) |
| 171 |
{ |
| 172 |
int tmp = 0; |
| 173 |
|
| 174 |
if (b == 0) |
| 175 |
return 0xffffffff; |
| 176 |
#ifdef LZ |
| 177 |
tmp = LZ (b); |
| 178 |
#else |
| 179 |
for (tmp = 31; (((long) b & (1 << tmp)) == 0); tmp--) |
| 180 |
; |
| 181 |
|
| 182 |
tmp = 31 - tmp; |
| 183 |
#endif |
| 184 |
|
| 185 |
if ((b << tmp) > a) |
| 186 |
{ |
| 187 |
extra = 1 << (tmp-1); |
| 188 |
a -= b << (tmp - 1); |
| 189 |
} |
| 190 |
else |
| 191 |
{ |
| 192 |
extra = 1 << tmp; |
| 193 |
a -= b << tmp; |
| 194 |
} |
| 195 |
} |
| 196 |
|
| 197 |
return do_31div (a, b).quot+extra; |
| 198 |
} |
| 199 |
#endif /* L_udivsi3 */ |
| 200 |
|
| 201 |
#ifdef L_divsi3 |
| 202 |
long |
| 203 |
__Div (long a, long b) __attribute__ ((__const__)); |
| 204 |
|
| 205 |
long |
| 206 |
__Div (long a, long b) |
| 207 |
{ |
| 208 |
long extra = 0; |
| 209 |
long sign = (b < 0) ? -1 : 1; |
| 210 |
long res; |
| 211 |
|
| 212 |
/* We need to handle a == -2147483648 as expected and must while |
| 213 |
doing that avoid producing a sequence like "abs (a) < 0" as GCC |
| 214 |
may optimize out the test. That sequence may not be obvious as |
| 215 |
we call inline functions. Testing for a being negative and |
| 216 |
handling (presumably much rarer than positive) enables us to get |
| 217 |
a bit of optimization for an (accumulated) reduction of the |
| 218 |
penalty of the 0x80000000 special-case. */ |
| 219 |
if (a < 0) |
| 220 |
{ |
| 221 |
sign = -sign; |
| 222 |
|
| 223 |
if ((a & 0x7fffffff) == 0) |
| 224 |
{ |
| 225 |
/* We're at 0x80000000. Tread carefully. */ |
| 226 |
a -= SIGNMULT (sign, b); |
| 227 |
extra = sign; |
| 228 |
} |
| 229 |
a = -a; |
| 230 |
} |
| 231 |
|
| 232 |
res = do_31div (a, __builtin_labs (b)).quot; |
| 233 |
return SIGNMULT (sign, res) + extra; |
| 234 |
} |
| 235 |
#endif /* L_divsi3 */ |
| 236 |
|
| 237 |
|
| 238 |
#ifdef L_umodsi3 |
| 239 |
unsigned long |
| 240 |
__Umod (unsigned long a, unsigned long b) __attribute__ ((__const__)); |
| 241 |
|
| 242 |
unsigned long |
| 243 |
__Umod (unsigned long a, unsigned long b) |
| 244 |
{ |
| 245 |
/* Adjust operands and result if a and/or b is 32 bits. */ |
| 246 |
if ((long) b < 0) |
| 247 |
return a >= b ? a - b : a; |
| 248 |
|
| 249 |
if ((long) a < 0) |
| 250 |
{ |
| 251 |
int tmp = 0; |
| 252 |
|
| 253 |
if (b == 0) |
| 254 |
return a; |
| 255 |
#ifdef LZ |
| 256 |
tmp = LZ (b); |
| 257 |
#else |
| 258 |
for (tmp = 31; (((long) b & (1 << tmp)) == 0); tmp--) |
| 259 |
; |
| 260 |
tmp = 31 - tmp; |
| 261 |
#endif |
| 262 |
|
| 263 |
if ((b << tmp) > a) |
| 264 |
{ |
| 265 |
a -= b << (tmp - 1); |
| 266 |
} |
| 267 |
else |
| 268 |
{ |
| 269 |
a -= b << tmp; |
| 270 |
} |
| 271 |
} |
| 272 |
|
| 273 |
return do_31div (a, b).rem; |
| 274 |
} |
| 275 |
#endif /* L_umodsi3 */ |
| 276 |
|
| 277 |
#ifdef L_modsi3 |
| 278 |
long |
| 279 |
__Mod (long a, long b) __attribute__ ((__const__)); |
| 280 |
|
| 281 |
long |
| 282 |
__Mod (long a, long b) |
| 283 |
{ |
| 284 |
long sign = 1; |
| 285 |
long res; |
| 286 |
|
| 287 |
/* We need to handle a == -2147483648 as expected and must while |
| 288 |
doing that avoid producing a sequence like "abs (a) < 0" as GCC |
| 289 |
may optimize out the test. That sequence may not be obvious as |
| 290 |
we call inline functions. Testing for a being negative and |
| 291 |
handling (presumably much rarer than positive) enables us to get |
| 292 |
a bit of optimization for an (accumulated) reduction of the |
| 293 |
penalty of the 0x80000000 special-case. */ |
| 294 |
if (a < 0) |
| 295 |
{ |
| 296 |
sign = -1; |
| 297 |
if ((a & 0x7fffffff) == 0) |
| 298 |
/* We're at 0x80000000. Tread carefully. */ |
| 299 |
a += __builtin_labs (b); |
| 300 |
a = -a; |
| 301 |
} |
| 302 |
|
| 303 |
res = do_31div (a, __builtin_labs (b)).rem; |
| 304 |
return SIGNMULT (sign, res); |
| 305 |
} |
| 306 |
#endif /* L_modsi3 */ |
| 307 |
#endif /* L_udivsi3 || L_divsi3 || L_umodsi3 || L_modsi3 */ |
| 308 |
|
| 309 |
/* |
| 310 |
* Local variables: |
| 311 |
* eval: (c-set-style "gnu") |
| 312 |
* indent-tabs-mode: t |
| 313 |
* End: |
| 314 |
*/ |