| 1 |
/* This is an assembly language implementation of mulsi3, divsi3, and modsi3 |
| 2 |
for the sparc processor. |
| 3 |
|
| 4 |
These routines are derived from the SPARC Architecture Manual, version 8, |
| 5 |
slightly edited to match the desired calling convention, and also to |
| 6 |
optimize them for our purposes. */ |
| 7 |
|
| 8 |
/* An executable stack is *not* required for these functions. */ |
| 9 |
#if defined(__ELF__) && defined(__linux__) |
| 10 |
.section .note.GNU-stack,"",%progbits |
| 11 |
.previous |
| 12 |
#endif |
| 13 |
|
| 14 |
#ifdef L_mulsi3 |
| 15 |
.text |
| 16 |
.align 4 |
| 17 |
.global .umul |
| 18 |
.proc 4 |
| 19 |
.umul: |
| 20 |
or %o0, %o1, %o4 ! logical or of multiplier and multiplicand |
| 21 |
mov %o0, %y ! multiplier to Y register |
| 22 |
andncc %o4, 0xfff, %o5 ! mask out lower 12 bits |
| 23 |
be mul_shortway ! can do it the short way |
| 24 |
andcc %g0, %g0, %o4 ! zero the partial product and clear NV cc |
| 25 |
! |
| 26 |
! long multiply |
| 27 |
! |
| 28 |
mulscc %o4, %o1, %o4 ! first iteration of 33 |
| 29 |
mulscc %o4, %o1, %o4 |
| 30 |
mulscc %o4, %o1, %o4 |
| 31 |
mulscc %o4, %o1, %o4 |
| 32 |
mulscc %o4, %o1, %o4 |
| 33 |
mulscc %o4, %o1, %o4 |
| 34 |
mulscc %o4, %o1, %o4 |
| 35 |
mulscc %o4, %o1, %o4 |
| 36 |
mulscc %o4, %o1, %o4 |
| 37 |
mulscc %o4, %o1, %o4 |
| 38 |
mulscc %o4, %o1, %o4 |
| 39 |
mulscc %o4, %o1, %o4 |
| 40 |
mulscc %o4, %o1, %o4 |
| 41 |
mulscc %o4, %o1, %o4 |
| 42 |
mulscc %o4, %o1, %o4 |
| 43 |
mulscc %o4, %o1, %o4 |
| 44 |
mulscc %o4, %o1, %o4 |
| 45 |
mulscc %o4, %o1, %o4 |
| 46 |
mulscc %o4, %o1, %o4 |
| 47 |
mulscc %o4, %o1, %o4 |
| 48 |
mulscc %o4, %o1, %o4 |
| 49 |
mulscc %o4, %o1, %o4 |
| 50 |
mulscc %o4, %o1, %o4 |
| 51 |
mulscc %o4, %o1, %o4 |
| 52 |
mulscc %o4, %o1, %o4 |
| 53 |
mulscc %o4, %o1, %o4 |
| 54 |
mulscc %o4, %o1, %o4 |
| 55 |
mulscc %o4, %o1, %o4 |
| 56 |
mulscc %o4, %o1, %o4 |
| 57 |
mulscc %o4, %o1, %o4 |
| 58 |
mulscc %o4, %o1, %o4 |
| 59 |
mulscc %o4, %o1, %o4 ! 32nd iteration |
| 60 |
mulscc %o4, %g0, %o4 ! last iteration only shifts |
| 61 |
! the upper 32 bits of product are wrong, but we do not care |
| 62 |
retl |
| 63 |
rd %y, %o0 |
| 64 |
! |
| 65 |
! short multiply |
| 66 |
! |
| 67 |
mul_shortway: |
| 68 |
mulscc %o4, %o1, %o4 ! first iteration of 13 |
| 69 |
mulscc %o4, %o1, %o4 |
| 70 |
mulscc %o4, %o1, %o4 |
| 71 |
mulscc %o4, %o1, %o4 |
| 72 |
mulscc %o4, %o1, %o4 |
| 73 |
mulscc %o4, %o1, %o4 |
| 74 |
mulscc %o4, %o1, %o4 |
| 75 |
mulscc %o4, %o1, %o4 |
| 76 |
mulscc %o4, %o1, %o4 |
| 77 |
mulscc %o4, %o1, %o4 |
| 78 |
mulscc %o4, %o1, %o4 |
| 79 |
mulscc %o4, %o1, %o4 ! 12th iteration |
| 80 |
mulscc %o4, %g0, %o4 ! last iteration only shifts |
| 81 |
rd %y, %o5 |
| 82 |
sll %o4, 12, %o4 ! left shift partial product by 12 bits |
| 83 |
srl %o5, 20, %o5 ! right shift partial product by 20 bits |
| 84 |
retl |
| 85 |
or %o5, %o4, %o0 ! merge for true product |
| 86 |
#endif |
| 87 |
|
| 88 |
#ifdef L_divsi3 |
| 89 |
/* |
| 90 |
* Division and remainder, from Appendix E of the SPARC Version 8 |
| 91 |
* Architecture Manual, with fixes from Gordon Irlam. |
| 92 |
*/ |
| 93 |
|
| 94 |
/* |
| 95 |
* Input: dividend and divisor in %o0 and %o1 respectively. |
| 96 |
* |
| 97 |
* m4 parameters: |
| 98 |
* .div name of function to generate |
| 99 |
* div div=div => %o0 / %o1; div=rem => %o0 % %o1 |
| 100 |
* true true=true => signed; true=false => unsigned |
| 101 |
* |
| 102 |
* Algorithm parameters: |
| 103 |
* N how many bits per iteration we try to get (4) |
| 104 |
* WORDSIZE total number of bits (32) |
| 105 |
* |
| 106 |
* Derived constants: |
| 107 |
* TOPBITS number of bits in the top decade of a number |
| 108 |
* |
| 109 |
* Important variables: |
| 110 |
* Q the partial quotient under development (initially 0) |
| 111 |
* R the remainder so far, initially the dividend |
| 112 |
* ITER number of main division loop iterations required; |
| 113 |
* equal to ceil(log2(quotient) / N). Note that this |
| 114 |
* is the log base (2^N) of the quotient. |
| 115 |
* V the current comparand, initially divisor*2^(ITER*N-1) |
| 116 |
* |
| 117 |
* Cost: |
| 118 |
* Current estimate for non-large dividend is |
| 119 |
* ceil(log2(quotient) / N) * (10 + 7N/2) + C |
| 120 |
* A large dividend is one greater than 2^(31-TOPBITS) and takes a |
| 121 |
* different path, as the upper bits of the quotient must be developed |
| 122 |
* one bit at a time. |
| 123 |
*/ |
| 124 |
.global .udiv |
| 125 |
.align 4 |
| 126 |
.proc 4 |
| 127 |
.text |
| 128 |
.udiv: |
| 129 |
b ready_to_divide |
| 130 |
mov 0, %g3 ! result is always positive |
| 131 |
|
| 132 |
.global .div |
| 133 |
.align 4 |
| 134 |
.proc 4 |
| 135 |
.text |
| 136 |
.div: |
| 137 |
! compute sign of result; if neither is negative, no problem |
| 138 |
orcc %o1, %o0, %g0 ! either negative? |
| 139 |
bge ready_to_divide ! no, go do the divide |
| 140 |
xor %o1, %o0, %g3 ! compute sign in any case |
| 141 |
tst %o1 |
| 142 |
bge 1f |
| 143 |
tst %o0 |
| 144 |
! %o1 is definitely negative; %o0 might also be negative |
| 145 |
bge ready_to_divide ! if %o0 not negative... |
| 146 |
sub %g0, %o1, %o1 ! in any case, make %o1 nonneg |
| 147 |
1: ! %o0 is negative, %o1 is nonnegative |
| 148 |
sub %g0, %o0, %o0 ! make %o0 nonnegative |
| 149 |
|
| 150 |
|
| 151 |
ready_to_divide: |
| 152 |
|
| 153 |
! Ready to divide. Compute size of quotient; scale comparand. |
| 154 |
orcc %o1, %g0, %o5 |
| 155 |
bne 1f |
| 156 |
mov %o0, %o3 |
| 157 |
|
| 158 |
! Divide by zero trap. If it returns, return 0 (about as |
| 159 |
! wrong as possible, but that is what SunOS does...). |
| 160 |
ta 0x2 ! ST_DIV0 |
| 161 |
retl |
| 162 |
clr %o0 |
| 163 |
|
| 164 |
1: |
| 165 |
cmp %o3, %o5 ! if %o1 exceeds %o0, done |
| 166 |
blu got_result ! (and algorithm fails otherwise) |
| 167 |
clr %o2 |
| 168 |
sethi %hi(1 << (32 - 4 - 1)), %g1 |
| 169 |
cmp %o3, %g1 |
| 170 |
blu not_really_big |
| 171 |
clr %o4 |
| 172 |
|
| 173 |
! Here the dividend is >= 2**(31-N) or so. We must be careful here, |
| 174 |
! as our usual N-at-a-shot divide step will cause overflow and havoc. |
| 175 |
! The number of bits in the result here is N*ITER+SC, where SC <= N. |
| 176 |
! Compute ITER in an unorthodox manner: know we need to shift V into |
| 177 |
! the top decade: so do not even bother to compare to R. |
| 178 |
1: |
| 179 |
cmp %o5, %g1 |
| 180 |
bgeu 3f |
| 181 |
mov 1, %g2 |
| 182 |
sll %o5, 4, %o5 |
| 183 |
b 1b |
| 184 |
add %o4, 1, %o4 |
| 185 |
|
| 186 |
! Now compute %g2. |
| 187 |
2: addcc %o5, %o5, %o5 |
| 188 |
bcc not_too_big |
| 189 |
add %g2, 1, %g2 |
| 190 |
|
| 191 |
! We get here if the %o1 overflowed while shifting. |
| 192 |
! This means that %o3 has the high-order bit set. |
| 193 |
! Restore %o5 and subtract from %o3. |
| 194 |
sll %g1, 4, %g1 ! high order bit |
| 195 |
srl %o5, 1, %o5 ! rest of %o5 |
| 196 |
add %o5, %g1, %o5 |
| 197 |
b do_single_div |
| 198 |
sub %g2, 1, %g2 |
| 199 |
|
| 200 |
not_too_big: |
| 201 |
3: cmp %o5, %o3 |
| 202 |
blu 2b |
| 203 |
nop |
| 204 |
be do_single_div |
| 205 |
nop |
| 206 |
/* NB: these are commented out in the V8-SPARC manual as well */ |
| 207 |
/* (I do not understand this) */ |
| 208 |
! %o5 > %o3: went too far: back up 1 step |
| 209 |
! srl %o5, 1, %o5 |
| 210 |
! dec %g2 |
| 211 |
! do single-bit divide steps |
| 212 |
! |
| 213 |
! We have to be careful here. We know that %o3 >= %o5, so we can do the |
| 214 |
! first divide step without thinking. BUT, the others are conditional, |
| 215 |
! and are only done if %o3 >= 0. Because both %o3 and %o5 may have the high- |
| 216 |
! order bit set in the first step, just falling into the regular |
| 217 |
! division loop will mess up the first time around. |
| 218 |
! So we unroll slightly... |
| 219 |
do_single_div: |
| 220 |
subcc %g2, 1, %g2 |
| 221 |
bl end_regular_divide |
| 222 |
nop |
| 223 |
sub %o3, %o5, %o3 |
| 224 |
mov 1, %o2 |
| 225 |
b end_single_divloop |
| 226 |
nop |
| 227 |
single_divloop: |
| 228 |
sll %o2, 1, %o2 |
| 229 |
bl 1f |
| 230 |
srl %o5, 1, %o5 |
| 231 |
! %o3 >= 0 |
| 232 |
sub %o3, %o5, %o3 |
| 233 |
b 2f |
| 234 |
add %o2, 1, %o2 |
| 235 |
1: ! %o3 < 0 |
| 236 |
add %o3, %o5, %o3 |
| 237 |
sub %o2, 1, %o2 |
| 238 |
2: |
| 239 |
end_single_divloop: |
| 240 |
subcc %g2, 1, %g2 |
| 241 |
bge single_divloop |
| 242 |
tst %o3 |
| 243 |
b,a end_regular_divide |
| 244 |
|
| 245 |
not_really_big: |
| 246 |
1: |
| 247 |
sll %o5, 4, %o5 |
| 248 |
cmp %o5, %o3 |
| 249 |
bleu 1b |
| 250 |
addcc %o4, 1, %o4 |
| 251 |
be got_result |
| 252 |
sub %o4, 1, %o4 |
| 253 |
|
| 254 |
tst %o3 ! set up for initial iteration |
| 255 |
divloop: |
| 256 |
sll %o2, 4, %o2 |
| 257 |
! depth 1, accumulated bits 0 |
| 258 |
bl L1.16 |
| 259 |
srl %o5,1,%o5 |
| 260 |
! remainder is positive |
| 261 |
subcc %o3,%o5,%o3 |
| 262 |
! depth 2, accumulated bits 1 |
| 263 |
bl L2.17 |
| 264 |
srl %o5,1,%o5 |
| 265 |
! remainder is positive |
| 266 |
subcc %o3,%o5,%o3 |
| 267 |
! depth 3, accumulated bits 3 |
| 268 |
bl L3.19 |
| 269 |
srl %o5,1,%o5 |
| 270 |
! remainder is positive |
| 271 |
subcc %o3,%o5,%o3 |
| 272 |
! depth 4, accumulated bits 7 |
| 273 |
bl L4.23 |
| 274 |
srl %o5,1,%o5 |
| 275 |
! remainder is positive |
| 276 |
subcc %o3,%o5,%o3 |
| 277 |
b 9f |
| 278 |
add %o2, (7*2+1), %o2 |
| 279 |
|
| 280 |
L4.23: |
| 281 |
! remainder is negative |
| 282 |
addcc %o3,%o5,%o3 |
| 283 |
b 9f |
| 284 |
add %o2, (7*2-1), %o2 |
| 285 |
|
| 286 |
|
| 287 |
L3.19: |
| 288 |
! remainder is negative |
| 289 |
addcc %o3,%o5,%o3 |
| 290 |
! depth 4, accumulated bits 5 |
| 291 |
bl L4.21 |
| 292 |
srl %o5,1,%o5 |
| 293 |
! remainder is positive |
| 294 |
subcc %o3,%o5,%o3 |
| 295 |
b 9f |
| 296 |
add %o2, (5*2+1), %o2 |
| 297 |
|
| 298 |
L4.21: |
| 299 |
! remainder is negative |
| 300 |
addcc %o3,%o5,%o3 |
| 301 |
b 9f |
| 302 |
add %o2, (5*2-1), %o2 |
| 303 |
|
| 304 |
L2.17: |
| 305 |
! remainder is negative |
| 306 |
addcc %o3,%o5,%o3 |
| 307 |
! depth 3, accumulated bits 1 |
| 308 |
bl L3.17 |
| 309 |
srl %o5,1,%o5 |
| 310 |
! remainder is positive |
| 311 |
subcc %o3,%o5,%o3 |
| 312 |
! depth 4, accumulated bits 3 |
| 313 |
bl L4.19 |
| 314 |
srl %o5,1,%o5 |
| 315 |
! remainder is positive |
| 316 |
subcc %o3,%o5,%o3 |
| 317 |
b 9f |
| 318 |
add %o2, (3*2+1), %o2 |
| 319 |
|
| 320 |
L4.19: |
| 321 |
! remainder is negative |
| 322 |
addcc %o3,%o5,%o3 |
| 323 |
b 9f |
| 324 |
add %o2, (3*2-1), %o2 |
| 325 |
|
| 326 |
L3.17: |
| 327 |
! remainder is negative |
| 328 |
addcc %o3,%o5,%o3 |
| 329 |
! depth 4, accumulated bits 1 |
| 330 |
bl L4.17 |
| 331 |
srl %o5,1,%o5 |
| 332 |
! remainder is positive |
| 333 |
subcc %o3,%o5,%o3 |
| 334 |
b 9f |
| 335 |
add %o2, (1*2+1), %o2 |
| 336 |
|
| 337 |
L4.17: |
| 338 |
! remainder is negative |
| 339 |
addcc %o3,%o5,%o3 |
| 340 |
b 9f |
| 341 |
add %o2, (1*2-1), %o2 |
| 342 |
|
| 343 |
L1.16: |
| 344 |
! remainder is negative |
| 345 |
addcc %o3,%o5,%o3 |
| 346 |
! depth 2, accumulated bits -1 |
| 347 |
bl L2.15 |
| 348 |
srl %o5,1,%o5 |
| 349 |
! remainder is positive |
| 350 |
subcc %o3,%o5,%o3 |
| 351 |
! depth 3, accumulated bits -1 |
| 352 |
bl L3.15 |
| 353 |
srl %o5,1,%o5 |
| 354 |
! remainder is positive |
| 355 |
subcc %o3,%o5,%o3 |
| 356 |
! depth 4, accumulated bits -1 |
| 357 |
bl L4.15 |
| 358 |
srl %o5,1,%o5 |
| 359 |
! remainder is positive |
| 360 |
subcc %o3,%o5,%o3 |
| 361 |
b 9f |
| 362 |
add %o2, (-1*2+1), %o2 |
| 363 |
|
| 364 |
L4.15: |
| 365 |
! remainder is negative |
| 366 |
addcc %o3,%o5,%o3 |
| 367 |
b 9f |
| 368 |
add %o2, (-1*2-1), %o2 |
| 369 |
|
| 370 |
L3.15: |
| 371 |
! remainder is negative |
| 372 |
addcc %o3,%o5,%o3 |
| 373 |
! depth 4, accumulated bits -3 |
| 374 |
bl L4.13 |
| 375 |
srl %o5,1,%o5 |
| 376 |
! remainder is positive |
| 377 |
subcc %o3,%o5,%o3 |
| 378 |
b 9f |
| 379 |
add %o2, (-3*2+1), %o2 |
| 380 |
|
| 381 |
L4.13: |
| 382 |
! remainder is negative |
| 383 |
addcc %o3,%o5,%o3 |
| 384 |
b 9f |
| 385 |
add %o2, (-3*2-1), %o2 |
| 386 |
|
| 387 |
L2.15: |
| 388 |
! remainder is negative |
| 389 |
addcc %o3,%o5,%o3 |
| 390 |
! depth 3, accumulated bits -3 |
| 391 |
bl L3.13 |
| 392 |
srl %o5,1,%o5 |
| 393 |
! remainder is positive |
| 394 |
subcc %o3,%o5,%o3 |
| 395 |
! depth 4, accumulated bits -5 |
| 396 |
bl L4.11 |
| 397 |
srl %o5,1,%o5 |
| 398 |
! remainder is positive |
| 399 |
subcc %o3,%o5,%o3 |
| 400 |
b 9f |
| 401 |
add %o2, (-5*2+1), %o2 |
| 402 |
|
| 403 |
L4.11: |
| 404 |
! remainder is negative |
| 405 |
addcc %o3,%o5,%o3 |
| 406 |
b 9f |
| 407 |
add %o2, (-5*2-1), %o2 |
| 408 |
|
| 409 |
L3.13: |
| 410 |
! remainder is negative |
| 411 |
addcc %o3,%o5,%o3 |
| 412 |
! depth 4, accumulated bits -7 |
| 413 |
bl L4.9 |
| 414 |
srl %o5,1,%o5 |
| 415 |
! remainder is positive |
| 416 |
subcc %o3,%o5,%o3 |
| 417 |
b 9f |
| 418 |
add %o2, (-7*2+1), %o2 |
| 419 |
|
| 420 |
L4.9: |
| 421 |
! remainder is negative |
| 422 |
addcc %o3,%o5,%o3 |
| 423 |
b 9f |
| 424 |
add %o2, (-7*2-1), %o2 |
| 425 |
|
| 426 |
9: |
| 427 |
end_regular_divide: |
| 428 |
subcc %o4, 1, %o4 |
| 429 |
bge divloop |
| 430 |
tst %o3 |
| 431 |
bl,a got_result |
| 432 |
! non-restoring fixup here (one instruction only!) |
| 433 |
sub %o2, 1, %o2 |
| 434 |
|
| 435 |
|
| 436 |
got_result: |
| 437 |
! check to see if answer should be < 0 |
| 438 |
tst %g3 |
| 439 |
bl,a 1f |
| 440 |
sub %g0, %o2, %o2 |
| 441 |
1: |
| 442 |
retl |
| 443 |
mov %o2, %o0 |
| 444 |
#endif |
| 445 |
|
| 446 |
#ifdef L_modsi3 |
| 447 |
/* This implementation was taken from glibc: |
| 448 |
* |
| 449 |
* Input: dividend and divisor in %o0 and %o1 respectively. |
| 450 |
* |
| 451 |
* Algorithm parameters: |
| 452 |
* N how many bits per iteration we try to get (4) |
| 453 |
* WORDSIZE total number of bits (32) |
| 454 |
* |
| 455 |
* Derived constants: |
| 456 |
* TOPBITS number of bits in the top decade of a number |
| 457 |
* |
| 458 |
* Important variables: |
| 459 |
* Q the partial quotient under development (initially 0) |
| 460 |
* R the remainder so far, initially the dividend |
| 461 |
* ITER number of main division loop iterations required; |
| 462 |
* equal to ceil(log2(quotient) / N). Note that this |
| 463 |
* is the log base (2^N) of the quotient. |
| 464 |
* V the current comparand, initially divisor*2^(ITER*N-1) |
| 465 |
* |
| 466 |
* Cost: |
| 467 |
* Current estimate for non-large dividend is |
| 468 |
* ceil(log2(quotient) / N) * (10 + 7N/2) + C |
| 469 |
* A large dividend is one greater than 2^(31-TOPBITS) and takes a |
| 470 |
* different path, as the upper bits of the quotient must be developed |
| 471 |
* one bit at a time. |
| 472 |
*/ |
| 473 |
.text |
| 474 |
.align 4 |
| 475 |
.global .urem |
| 476 |
.proc 4 |
| 477 |
.urem: |
| 478 |
b divide |
| 479 |
mov 0, %g3 ! result always positive |
| 480 |
|
| 481 |
.align 4 |
| 482 |
.global .rem |
| 483 |
.proc 4 |
| 484 |
.rem: |
| 485 |
! compute sign of result; if neither is negative, no problem |
| 486 |
orcc %o1, %o0, %g0 ! either negative? |
| 487 |
bge 2f ! no, go do the divide |
| 488 |
mov %o0, %g3 ! sign of remainder matches %o0 |
| 489 |
tst %o1 |
| 490 |
bge 1f |
| 491 |
tst %o0 |
| 492 |
! %o1 is definitely negative; %o0 might also be negative |
| 493 |
bge 2f ! if %o0 not negative... |
| 494 |
sub %g0, %o1, %o1 ! in any case, make %o1 nonneg |
| 495 |
1: ! %o0 is negative, %o1 is nonnegative |
| 496 |
sub %g0, %o0, %o0 ! make %o0 nonnegative |
| 497 |
2: |
| 498 |
|
| 499 |
! Ready to divide. Compute size of quotient; scale comparand. |
| 500 |
divide: |
| 501 |
orcc %o1, %g0, %o5 |
| 502 |
bne 1f |
| 503 |
mov %o0, %o3 |
| 504 |
|
| 505 |
! Divide by zero trap. If it returns, return 0 (about as |
| 506 |
! wrong as possible, but that is what SunOS does...). |
| 507 |
ta 0x2 !ST_DIV0 |
| 508 |
retl |
| 509 |
clr %o0 |
| 510 |
|
| 511 |
1: |
| 512 |
cmp %o3, %o5 ! if %o1 exceeds %o0, done |
| 513 |
blu got_result ! (and algorithm fails otherwise) |
| 514 |
clr %o2 |
| 515 |
sethi %hi(1 << (32 - 4 - 1)), %g1 |
| 516 |
cmp %o3, %g1 |
| 517 |
blu not_really_big |
| 518 |
clr %o4 |
| 519 |
|
| 520 |
! Here the dividend is >= 2**(31-N) or so. We must be careful here, |
| 521 |
! as our usual N-at-a-shot divide step will cause overflow and havoc. |
| 522 |
! The number of bits in the result here is N*ITER+SC, where SC <= N. |
| 523 |
! Compute ITER in an unorthodox manner: know we need to shift V into |
| 524 |
! the top decade: so do not even bother to compare to R. |
| 525 |
1: |
| 526 |
cmp %o5, %g1 |
| 527 |
bgeu 3f |
| 528 |
mov 1, %g2 |
| 529 |
sll %o5, 4, %o5 |
| 530 |
b 1b |
| 531 |
add %o4, 1, %o4 |
| 532 |
|
| 533 |
! Now compute %g2. |
| 534 |
2: addcc %o5, %o5, %o5 |
| 535 |
bcc not_too_big |
| 536 |
add %g2, 1, %g2 |
| 537 |
|
| 538 |
! We get here if the %o1 overflowed while shifting. |
| 539 |
! This means that %o3 has the high-order bit set. |
| 540 |
! Restore %o5 and subtract from %o3. |
| 541 |
sll %g1, 4, %g1 ! high order bit |
| 542 |
srl %o5, 1, %o5 ! rest of %o5 |
| 543 |
add %o5, %g1, %o5 |
| 544 |
b do_single_div |
| 545 |
sub %g2, 1, %g2 |
| 546 |
|
| 547 |
not_too_big: |
| 548 |
3: cmp %o5, %o3 |
| 549 |
blu 2b |
| 550 |
nop |
| 551 |
be do_single_div |
| 552 |
nop |
| 553 |
/* NB: these are commented out in the V8-SPARC manual as well */ |
| 554 |
/* (I do not understand this) */ |
| 555 |
! %o5 > %o3: went too far: back up 1 step |
| 556 |
! srl %o5, 1, %o5 |
| 557 |
! dec %g2 |
| 558 |
! do single-bit divide steps |
| 559 |
! |
| 560 |
! We have to be careful here. We know that %o3 >= %o5, so we can do the |
| 561 |
! first divide step without thinking. BUT, the others are conditional, |
| 562 |
! and are only done if %o3 >= 0. Because both %o3 and %o5 may have the high- |
| 563 |
! order bit set in the first step, just falling into the regular |
| 564 |
! division loop will mess up the first time around. |
| 565 |
! So we unroll slightly... |
| 566 |
do_single_div: |
| 567 |
subcc %g2, 1, %g2 |
| 568 |
bl end_regular_divide |
| 569 |
nop |
| 570 |
sub %o3, %o5, %o3 |
| 571 |
mov 1, %o2 |
| 572 |
b end_single_divloop |
| 573 |
nop |
| 574 |
single_divloop: |
| 575 |
sll %o2, 1, %o2 |
| 576 |
bl 1f |
| 577 |
srl %o5, 1, %o5 |
| 578 |
! %o3 >= 0 |
| 579 |
sub %o3, %o5, %o3 |
| 580 |
b 2f |
| 581 |
add %o2, 1, %o2 |
| 582 |
1: ! %o3 < 0 |
| 583 |
add %o3, %o5, %o3 |
| 584 |
sub %o2, 1, %o2 |
| 585 |
2: |
| 586 |
end_single_divloop: |
| 587 |
subcc %g2, 1, %g2 |
| 588 |
bge single_divloop |
| 589 |
tst %o3 |
| 590 |
b,a end_regular_divide |
| 591 |
|
| 592 |
not_really_big: |
| 593 |
1: |
| 594 |
sll %o5, 4, %o5 |
| 595 |
cmp %o5, %o3 |
| 596 |
bleu 1b |
| 597 |
addcc %o4, 1, %o4 |
| 598 |
be got_result |
| 599 |
sub %o4, 1, %o4 |
| 600 |
|
| 601 |
tst %o3 ! set up for initial iteration |
| 602 |
divloop: |
| 603 |
sll %o2, 4, %o2 |
| 604 |
! depth 1, accumulated bits 0 |
| 605 |
bl L1.16 |
| 606 |
srl %o5,1,%o5 |
| 607 |
! remainder is positive |
| 608 |
subcc %o3,%o5,%o3 |
| 609 |
! depth 2, accumulated bits 1 |
| 610 |
bl L2.17 |
| 611 |
srl %o5,1,%o5 |
| 612 |
! remainder is positive |
| 613 |
subcc %o3,%o5,%o3 |
| 614 |
! depth 3, accumulated bits 3 |
| 615 |
bl L3.19 |
| 616 |
srl %o5,1,%o5 |
| 617 |
! remainder is positive |
| 618 |
subcc %o3,%o5,%o3 |
| 619 |
! depth 4, accumulated bits 7 |
| 620 |
bl L4.23 |
| 621 |
srl %o5,1,%o5 |
| 622 |
! remainder is positive |
| 623 |
subcc %o3,%o5,%o3 |
| 624 |
b 9f |
| 625 |
add %o2, (7*2+1), %o2 |
| 626 |
L4.23: |
| 627 |
! remainder is negative |
| 628 |
addcc %o3,%o5,%o3 |
| 629 |
b 9f |
| 630 |
add %o2, (7*2-1), %o2 |
| 631 |
|
| 632 |
L3.19: |
| 633 |
! remainder is negative |
| 634 |
addcc %o3,%o5,%o3 |
| 635 |
! depth 4, accumulated bits 5 |
| 636 |
bl L4.21 |
| 637 |
srl %o5,1,%o5 |
| 638 |
! remainder is positive |
| 639 |
subcc %o3,%o5,%o3 |
| 640 |
b 9f |
| 641 |
add %o2, (5*2+1), %o2 |
| 642 |
|
| 643 |
L4.21: |
| 644 |
! remainder is negative |
| 645 |
addcc %o3,%o5,%o3 |
| 646 |
b 9f |
| 647 |
add %o2, (5*2-1), %o2 |
| 648 |
|
| 649 |
L2.17: |
| 650 |
! remainder is negative |
| 651 |
addcc %o3,%o5,%o3 |
| 652 |
! depth 3, accumulated bits 1 |
| 653 |
bl L3.17 |
| 654 |
srl %o5,1,%o5 |
| 655 |
! remainder is positive |
| 656 |
subcc %o3,%o5,%o3 |
| 657 |
! depth 4, accumulated bits 3 |
| 658 |
bl L4.19 |
| 659 |
srl %o5,1,%o5 |
| 660 |
! remainder is positive |
| 661 |
subcc %o3,%o5,%o3 |
| 662 |
b 9f |
| 663 |
add %o2, (3*2+1), %o2 |
| 664 |
|
| 665 |
L4.19: |
| 666 |
! remainder is negative |
| 667 |
addcc %o3,%o5,%o3 |
| 668 |
b 9f |
| 669 |
add %o2, (3*2-1), %o2 |
| 670 |
|
| 671 |
L3.17: |
| 672 |
! remainder is negative |
| 673 |
addcc %o3,%o5,%o3 |
| 674 |
! depth 4, accumulated bits 1 |
| 675 |
bl L4.17 |
| 676 |
srl %o5,1,%o5 |
| 677 |
! remainder is positive |
| 678 |
subcc %o3,%o5,%o3 |
| 679 |
b 9f |
| 680 |
add %o2, (1*2+1), %o2 |
| 681 |
|
| 682 |
L4.17: |
| 683 |
! remainder is negative |
| 684 |
addcc %o3,%o5,%o3 |
| 685 |
b 9f |
| 686 |
add %o2, (1*2-1), %o2 |
| 687 |
|
| 688 |
L1.16: |
| 689 |
! remainder is negative |
| 690 |
addcc %o3,%o5,%o3 |
| 691 |
! depth 2, accumulated bits -1 |
| 692 |
bl L2.15 |
| 693 |
srl %o5,1,%o5 |
| 694 |
! remainder is positive |
| 695 |
subcc %o3,%o5,%o3 |
| 696 |
! depth 3, accumulated bits -1 |
| 697 |
bl L3.15 |
| 698 |
srl %o5,1,%o5 |
| 699 |
! remainder is positive |
| 700 |
subcc %o3,%o5,%o3 |
| 701 |
! depth 4, accumulated bits -1 |
| 702 |
bl L4.15 |
| 703 |
srl %o5,1,%o5 |
| 704 |
! remainder is positive |
| 705 |
subcc %o3,%o5,%o3 |
| 706 |
b 9f |
| 707 |
add %o2, (-1*2+1), %o2 |
| 708 |
|
| 709 |
L4.15: |
| 710 |
! remainder is negative |
| 711 |
addcc %o3,%o5,%o3 |
| 712 |
b 9f |
| 713 |
add %o2, (-1*2-1), %o2 |
| 714 |
|
| 715 |
L3.15: |
| 716 |
! remainder is negative |
| 717 |
addcc %o3,%o5,%o3 |
| 718 |
! depth 4, accumulated bits -3 |
| 719 |
bl L4.13 |
| 720 |
srl %o5,1,%o5 |
| 721 |
! remainder is positive |
| 722 |
subcc %o3,%o5,%o3 |
| 723 |
b 9f |
| 724 |
add %o2, (-3*2+1), %o2 |
| 725 |
|
| 726 |
L4.13: |
| 727 |
! remainder is negative |
| 728 |
addcc %o3,%o5,%o3 |
| 729 |
b 9f |
| 730 |
add %o2, (-3*2-1), %o2 |
| 731 |
|
| 732 |
L2.15: |
| 733 |
! remainder is negative |
| 734 |
addcc %o3,%o5,%o3 |
| 735 |
! depth 3, accumulated bits -3 |
| 736 |
bl L3.13 |
| 737 |
srl %o5,1,%o5 |
| 738 |
! remainder is positive |
| 739 |
subcc %o3,%o5,%o3 |
| 740 |
! depth 4, accumulated bits -5 |
| 741 |
bl L4.11 |
| 742 |
srl %o5,1,%o5 |
| 743 |
! remainder is positive |
| 744 |
subcc %o3,%o5,%o3 |
| 745 |
b 9f |
| 746 |
add %o2, (-5*2+1), %o2 |
| 747 |
|
| 748 |
L4.11: |
| 749 |
! remainder is negative |
| 750 |
addcc %o3,%o5,%o3 |
| 751 |
b 9f |
| 752 |
add %o2, (-5*2-1), %o2 |
| 753 |
|
| 754 |
L3.13: |
| 755 |
! remainder is negative |
| 756 |
addcc %o3,%o5,%o3 |
| 757 |
! depth 4, accumulated bits -7 |
| 758 |
bl L4.9 |
| 759 |
srl %o5,1,%o5 |
| 760 |
! remainder is positive |
| 761 |
subcc %o3,%o5,%o3 |
| 762 |
b 9f |
| 763 |
add %o2, (-7*2+1), %o2 |
| 764 |
|
| 765 |
L4.9: |
| 766 |
! remainder is negative |
| 767 |
addcc %o3,%o5,%o3 |
| 768 |
b 9f |
| 769 |
add %o2, (-7*2-1), %o2 |
| 770 |
|
| 771 |
9: |
| 772 |
end_regular_divide: |
| 773 |
subcc %o4, 1, %o4 |
| 774 |
bge divloop |
| 775 |
tst %o3 |
| 776 |
bl,a got_result |
| 777 |
! non-restoring fixup here (one instruction only!) |
| 778 |
add %o3, %o1, %o3 |
| 779 |
|
| 780 |
got_result: |
| 781 |
! check to see if answer should be < 0 |
| 782 |
tst %g3 |
| 783 |
bl,a 1f |
| 784 |
sub %g0, %o3, %o3 |
| 785 |
1: |
| 786 |
retl |
| 787 |
mov %o3, %o0 |
| 788 |
|
| 789 |
#endif |
| 790 |
|