Greg Kroah-Hartman | b244131 | 2017-11-01 15:07:57 +0100 | [diff] [blame] | 1 | /* SPDX-License-Identifier: GPL-2.0 */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2 | .file "reg_u_div.S" |
| 3 | /*---------------------------------------------------------------------------+ |
| 4 | | reg_u_div.S | |
| 5 | | | |
| 6 | | Divide one FPU_REG by another and put the result in a destination FPU_REG.| |
| 7 | | | |
| 8 | | Copyright (C) 1992,1993,1995,1997 | |
| 9 | | W. Metzenthen, 22 Parker St, Ormond, Vic 3163, Australia | |
| 10 | | E-mail billm@suburbia.net | |
| 11 | | | |
| 12 | | | |
| 13 | +---------------------------------------------------------------------------*/ |
| 14 | |
| 15 | /*---------------------------------------------------------------------------+ |
| 16 | | Call from C as: | |
| 17 | | int FPU_u_div(FPU_REG *a, FPU_REG *b, FPU_REG *dest, | |
| 18 | | unsigned int control_word, char *sign) | |
| 19 | | | |
| 20 | | Does not compute the destination exponent, but does adjust it. | |
| 21 | | | |
| 22 | | Return value is the tag of the answer, or-ed with FPU_Exception if | |
| 23 | | one was raised, or -1 on internal error. | |
| 24 | +---------------------------------------------------------------------------*/ |
| 25 | |
| 26 | #include "exception.h" |
| 27 | #include "fpu_emu.h" |
| 28 | #include "control_w.h" |
| 29 | |
| 30 | |
| 31 | /* #define dSIGL(x) (x) */ |
| 32 | /* #define dSIGH(x) 4(x) */ |
| 33 | |
| 34 | |
| 35 | #ifndef NON_REENTRANT_FPU |
| 36 | /* |
| 37 | Local storage on the stack: |
| 38 | Result: FPU_accum_3:FPU_accum_2:FPU_accum_1:FPU_accum_0 |
| 39 | Overflow flag: ovfl_flag |
| 40 | */ |
| 41 | #define FPU_accum_3 -4(%ebp) |
| 42 | #define FPU_accum_2 -8(%ebp) |
| 43 | #define FPU_accum_1 -12(%ebp) |
| 44 | #define FPU_accum_0 -16(%ebp) |
| 45 | #define FPU_result_1 -20(%ebp) |
| 46 | #define FPU_result_2 -24(%ebp) |
| 47 | #define FPU_ovfl_flag -28(%ebp) |
| 48 | |
| 49 | #else |
| 50 | .data |
| 51 | /* |
| 52 | Local storage in a static area: |
| 53 | Result: FPU_accum_3:FPU_accum_2:FPU_accum_1:FPU_accum_0 |
| 54 | Overflow flag: ovfl_flag |
| 55 | */ |
| 56 | .align 4,0 |
| 57 | FPU_accum_3: |
| 58 | .long 0 |
| 59 | FPU_accum_2: |
| 60 | .long 0 |
| 61 | FPU_accum_1: |
| 62 | .long 0 |
| 63 | FPU_accum_0: |
| 64 | .long 0 |
| 65 | FPU_result_1: |
| 66 | .long 0 |
| 67 | FPU_result_2: |
| 68 | .long 0 |
| 69 | FPU_ovfl_flag: |
| 70 | .byte 0 |
| 71 | #endif /* NON_REENTRANT_FPU */ |
| 72 | |
| 73 | #define REGA PARAM1 |
| 74 | #define REGB PARAM2 |
| 75 | #define DEST PARAM3 |
| 76 | |
| 77 | .text |
Jiri Slaby | 6d685e5 | 2019-10-11 13:51:07 +0200 | [diff] [blame] | 78 | SYM_FUNC_START(FPU_u_div) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 79 | pushl %ebp |
| 80 | movl %esp,%ebp |
| 81 | #ifndef NON_REENTRANT_FPU |
| 82 | subl $28,%esp |
| 83 | #endif /* NON_REENTRANT_FPU */ |
| 84 | |
| 85 | pushl %esi |
| 86 | pushl %edi |
| 87 | pushl %ebx |
| 88 | |
| 89 | movl REGA,%esi |
| 90 | movl REGB,%ebx |
| 91 | movl DEST,%edi |
| 92 | |
| 93 | movswl EXP(%esi),%edx |
| 94 | movswl EXP(%ebx),%eax |
| 95 | subl %eax,%edx |
| 96 | addl EXP_BIAS,%edx |
| 97 | |
| 98 | /* A denormal and a large number can cause an exponent underflow */ |
| 99 | cmpl EXP_WAY_UNDER,%edx |
| 100 | jg xExp_not_underflow |
| 101 | |
| 102 | /* Set to a really low value allow correct handling */ |
| 103 | movl EXP_WAY_UNDER,%edx |
| 104 | |
| 105 | xExp_not_underflow: |
| 106 | |
| 107 | movw %dx,EXP(%edi) |
| 108 | |
| 109 | #ifdef PARANOID |
| 110 | /* testl $0x80000000, SIGH(%esi) // Dividend */ |
| 111 | /* je L_bugged */ |
| 112 | testl $0x80000000, SIGH(%ebx) /* Divisor */ |
| 113 | je L_bugged |
| 114 | #endif /* PARANOID */ |
| 115 | |
| 116 | /* Check if the divisor can be treated as having just 32 bits */ |
| 117 | cmpl $0,SIGL(%ebx) |
| 118 | jnz L_Full_Division /* Can't do a quick divide */ |
| 119 | |
| 120 | /* We should be able to zip through the division here */ |
| 121 | movl SIGH(%ebx),%ecx /* The divisor */ |
| 122 | movl SIGH(%esi),%edx /* Dividend */ |
| 123 | movl SIGL(%esi),%eax /* Dividend */ |
| 124 | |
| 125 | cmpl %ecx,%edx |
| 126 | setaeb FPU_ovfl_flag /* Keep a record */ |
| 127 | jb L_no_adjust |
| 128 | |
| 129 | subl %ecx,%edx /* Prevent the overflow */ |
| 130 | |
| 131 | L_no_adjust: |
| 132 | /* Divide the 64 bit number by the 32 bit denominator */ |
| 133 | divl %ecx |
| 134 | movl %eax,FPU_result_2 |
| 135 | |
| 136 | /* Work on the remainder of the first division */ |
| 137 | xorl %eax,%eax |
| 138 | divl %ecx |
| 139 | movl %eax,FPU_result_1 |
| 140 | |
| 141 | /* Work on the remainder of the 64 bit division */ |
| 142 | xorl %eax,%eax |
| 143 | divl %ecx |
| 144 | |
| 145 | testb $255,FPU_ovfl_flag /* was the num > denom ? */ |
| 146 | je L_no_overflow |
| 147 | |
| 148 | /* Do the shifting here */ |
| 149 | /* increase the exponent */ |
| 150 | incw EXP(%edi) |
| 151 | |
| 152 | /* shift the mantissa right one bit */ |
| 153 | stc /* To set the ms bit */ |
| 154 | rcrl FPU_result_2 |
| 155 | rcrl FPU_result_1 |
| 156 | rcrl %eax |
| 157 | |
| 158 | L_no_overflow: |
| 159 | jmp LRound_precision /* Do the rounding as required */ |
| 160 | |
| 161 | |
| 162 | /*---------------------------------------------------------------------------+ |
| 163 | | Divide: Return arg1/arg2 to arg3. | |
| 164 | | | |
| 165 | | This routine does not use the exponents of arg1 and arg2, but does | |
| 166 | | adjust the exponent of arg3. | |
| 167 | | | |
| 168 | | The maximum returned value is (ignoring exponents) | |
| 169 | | .ffffffff ffffffff | |
| 170 | | ------------------ = 1.ffffffff fffffffe | |
| 171 | | .80000000 00000000 | |
| 172 | | and the minimum is | |
| 173 | | .80000000 00000000 | |
| 174 | | ------------------ = .80000000 00000001 (rounded) | |
| 175 | | .ffffffff ffffffff | |
| 176 | | | |
| 177 | +---------------------------------------------------------------------------*/ |
| 178 | |
| 179 | |
| 180 | L_Full_Division: |
| 181 | /* Save extended dividend in local register */ |
| 182 | movl SIGL(%esi),%eax |
| 183 | movl %eax,FPU_accum_2 |
| 184 | movl SIGH(%esi),%eax |
| 185 | movl %eax,FPU_accum_3 |
| 186 | xorl %eax,%eax |
| 187 | movl %eax,FPU_accum_1 /* zero the extension */ |
| 188 | movl %eax,FPU_accum_0 /* zero the extension */ |
| 189 | |
| 190 | movl SIGL(%esi),%eax /* Get the current num */ |
| 191 | movl SIGH(%esi),%edx |
| 192 | |
| 193 | /*----------------------------------------------------------------------*/ |
| 194 | /* Initialization done. |
| 195 | Do the first 32 bits. */ |
| 196 | |
| 197 | movb $0,FPU_ovfl_flag |
| 198 | cmpl SIGH(%ebx),%edx /* Test for imminent overflow */ |
| 199 | jb LLess_than_1 |
| 200 | ja LGreater_than_1 |
| 201 | |
| 202 | cmpl SIGL(%ebx),%eax |
| 203 | jb LLess_than_1 |
| 204 | |
| 205 | LGreater_than_1: |
| 206 | /* The dividend is greater or equal, would cause overflow */ |
| 207 | setaeb FPU_ovfl_flag /* Keep a record */ |
| 208 | |
| 209 | subl SIGL(%ebx),%eax |
| 210 | sbbl SIGH(%ebx),%edx /* Prevent the overflow */ |
| 211 | movl %eax,FPU_accum_2 |
| 212 | movl %edx,FPU_accum_3 |
| 213 | |
| 214 | LLess_than_1: |
| 215 | /* At this point, we have a dividend < divisor, with a record of |
| 216 | adjustment in FPU_ovfl_flag */ |
| 217 | |
| 218 | /* We will divide by a number which is too large */ |
| 219 | movl SIGH(%ebx),%ecx |
| 220 | addl $1,%ecx |
| 221 | jnc LFirst_div_not_1 |
| 222 | |
| 223 | /* here we need to divide by 100000000h, |
| 224 | i.e., no division at all.. */ |
| 225 | mov %edx,%eax |
| 226 | jmp LFirst_div_done |
| 227 | |
| 228 | LFirst_div_not_1: |
| 229 | divl %ecx /* Divide the numerator by the augmented |
| 230 | denom ms dw */ |
| 231 | |
| 232 | LFirst_div_done: |
| 233 | movl %eax,FPU_result_2 /* Put the result in the answer */ |
| 234 | |
| 235 | mull SIGH(%ebx) /* mul by the ms dw of the denom */ |
| 236 | |
| 237 | subl %eax,FPU_accum_2 /* Subtract from the num local reg */ |
| 238 | sbbl %edx,FPU_accum_3 |
| 239 | |
| 240 | movl FPU_result_2,%eax /* Get the result back */ |
| 241 | mull SIGL(%ebx) /* now mul the ls dw of the denom */ |
| 242 | |
| 243 | subl %eax,FPU_accum_1 /* Subtract from the num local reg */ |
| 244 | sbbl %edx,FPU_accum_2 |
| 245 | sbbl $0,FPU_accum_3 |
| 246 | je LDo_2nd_32_bits /* Must check for non-zero result here */ |
| 247 | |
| 248 | #ifdef PARANOID |
| 249 | jb L_bugged_1 |
| 250 | #endif /* PARANOID */ |
| 251 | |
| 252 | /* need to subtract another once of the denom */ |
| 253 | incl FPU_result_2 /* Correct the answer */ |
| 254 | |
| 255 | movl SIGL(%ebx),%eax |
| 256 | movl SIGH(%ebx),%edx |
| 257 | subl %eax,FPU_accum_1 /* Subtract from the num local reg */ |
| 258 | sbbl %edx,FPU_accum_2 |
| 259 | |
| 260 | #ifdef PARANOID |
| 261 | sbbl $0,FPU_accum_3 |
| 262 | jne L_bugged_1 /* Must check for non-zero result here */ |
| 263 | #endif /* PARANOID */ |
| 264 | |
| 265 | /*----------------------------------------------------------------------*/ |
| 266 | /* Half of the main problem is done, there is just a reduced numerator |
| 267 | to handle now. |
| 268 | Work with the second 32 bits, FPU_accum_0 not used from now on */ |
| 269 | LDo_2nd_32_bits: |
| 270 | movl FPU_accum_2,%edx /* get the reduced num */ |
| 271 | movl FPU_accum_1,%eax |
| 272 | |
| 273 | /* need to check for possible subsequent overflow */ |
| 274 | cmpl SIGH(%ebx),%edx |
| 275 | jb LDo_2nd_div |
| 276 | ja LPrevent_2nd_overflow |
| 277 | |
| 278 | cmpl SIGL(%ebx),%eax |
| 279 | jb LDo_2nd_div |
| 280 | |
| 281 | LPrevent_2nd_overflow: |
| 282 | /* The numerator is greater or equal, would cause overflow */ |
| 283 | /* prevent overflow */ |
| 284 | subl SIGL(%ebx),%eax |
| 285 | sbbl SIGH(%ebx),%edx |
| 286 | movl %edx,FPU_accum_2 |
| 287 | movl %eax,FPU_accum_1 |
| 288 | |
| 289 | incl FPU_result_2 /* Reflect the subtraction in the answer */ |
| 290 | |
| 291 | #ifdef PARANOID |
| 292 | je L_bugged_2 /* Can't bump the result to 1.0 */ |
| 293 | #endif /* PARANOID */ |
| 294 | |
| 295 | LDo_2nd_div: |
| 296 | cmpl $0,%ecx /* augmented denom msw */ |
| 297 | jnz LSecond_div_not_1 |
| 298 | |
| 299 | /* %ecx == 0, we are dividing by 1.0 */ |
| 300 | mov %edx,%eax |
| 301 | jmp LSecond_div_done |
| 302 | |
| 303 | LSecond_div_not_1: |
| 304 | divl %ecx /* Divide the numerator by the denom ms dw */ |
| 305 | |
| 306 | LSecond_div_done: |
| 307 | movl %eax,FPU_result_1 /* Put the result in the answer */ |
| 308 | |
| 309 | mull SIGH(%ebx) /* mul by the ms dw of the denom */ |
| 310 | |
| 311 | subl %eax,FPU_accum_1 /* Subtract from the num local reg */ |
| 312 | sbbl %edx,FPU_accum_2 |
| 313 | |
| 314 | #ifdef PARANOID |
| 315 | jc L_bugged_2 |
| 316 | #endif /* PARANOID */ |
| 317 | |
| 318 | movl FPU_result_1,%eax /* Get the result back */ |
| 319 | mull SIGL(%ebx) /* now mul the ls dw of the denom */ |
| 320 | |
| 321 | subl %eax,FPU_accum_0 /* Subtract from the num local reg */ |
| 322 | sbbl %edx,FPU_accum_1 /* Subtract from the num local reg */ |
| 323 | sbbl $0,FPU_accum_2 |
| 324 | |
| 325 | #ifdef PARANOID |
| 326 | jc L_bugged_2 |
| 327 | #endif /* PARANOID */ |
| 328 | |
| 329 | jz LDo_3rd_32_bits |
| 330 | |
| 331 | #ifdef PARANOID |
| 332 | cmpl $1,FPU_accum_2 |
| 333 | jne L_bugged_2 |
| 334 | #endif /* PARANOID */ |
| 335 | |
| 336 | /* need to subtract another once of the denom */ |
| 337 | movl SIGL(%ebx),%eax |
| 338 | movl SIGH(%ebx),%edx |
| 339 | subl %eax,FPU_accum_0 /* Subtract from the num local reg */ |
| 340 | sbbl %edx,FPU_accum_1 |
| 341 | sbbl $0,FPU_accum_2 |
| 342 | |
| 343 | #ifdef PARANOID |
| 344 | jc L_bugged_2 |
| 345 | jne L_bugged_2 |
| 346 | #endif /* PARANOID */ |
| 347 | |
| 348 | addl $1,FPU_result_1 /* Correct the answer */ |
| 349 | adcl $0,FPU_result_2 |
| 350 | |
| 351 | #ifdef PARANOID |
| 352 | jc L_bugged_2 /* Must check for non-zero result here */ |
| 353 | #endif /* PARANOID */ |
| 354 | |
| 355 | /*----------------------------------------------------------------------*/ |
| 356 | /* The division is essentially finished here, we just need to perform |
| 357 | tidying operations. |
| 358 | Deal with the 3rd 32 bits */ |
| 359 | LDo_3rd_32_bits: |
| 360 | movl FPU_accum_1,%edx /* get the reduced num */ |
| 361 | movl FPU_accum_0,%eax |
| 362 | |
| 363 | /* need to check for possible subsequent overflow */ |
| 364 | cmpl SIGH(%ebx),%edx /* denom */ |
| 365 | jb LRound_prep |
| 366 | ja LPrevent_3rd_overflow |
| 367 | |
| 368 | cmpl SIGL(%ebx),%eax /* denom */ |
| 369 | jb LRound_prep |
| 370 | |
| 371 | LPrevent_3rd_overflow: |
| 372 | /* prevent overflow */ |
| 373 | subl SIGL(%ebx),%eax |
| 374 | sbbl SIGH(%ebx),%edx |
| 375 | movl %edx,FPU_accum_1 |
| 376 | movl %eax,FPU_accum_0 |
| 377 | |
| 378 | addl $1,FPU_result_1 /* Reflect the subtraction in the answer */ |
| 379 | adcl $0,FPU_result_2 |
| 380 | jne LRound_prep |
| 381 | jnc LRound_prep |
| 382 | |
| 383 | /* This is a tricky spot, there is an overflow of the answer */ |
| 384 | movb $255,FPU_ovfl_flag /* Overflow -> 1.000 */ |
| 385 | |
| 386 | LRound_prep: |
| 387 | /* |
| 388 | * Prepare for rounding. |
| 389 | * To test for rounding, we just need to compare 2*accum with the |
| 390 | * denom. |
| 391 | */ |
| 392 | movl FPU_accum_0,%ecx |
| 393 | movl FPU_accum_1,%edx |
| 394 | movl %ecx,%eax |
| 395 | orl %edx,%eax |
| 396 | jz LRound_ovfl /* The accumulator contains zero. */ |
| 397 | |
| 398 | /* Multiply by 2 */ |
| 399 | clc |
| 400 | rcll $1,%ecx |
| 401 | rcll $1,%edx |
| 402 | jc LRound_large /* No need to compare, denom smaller */ |
| 403 | |
| 404 | subl SIGL(%ebx),%ecx |
| 405 | sbbl SIGH(%ebx),%edx |
| 406 | jnc LRound_not_small |
| 407 | |
| 408 | movl $0x70000000,%eax /* Denom was larger */ |
| 409 | jmp LRound_ovfl |
| 410 | |
| 411 | LRound_not_small: |
| 412 | jnz LRound_large |
| 413 | |
| 414 | movl $0x80000000,%eax /* Remainder was exactly 1/2 denom */ |
| 415 | jmp LRound_ovfl |
| 416 | |
| 417 | LRound_large: |
| 418 | movl $0xff000000,%eax /* Denom was smaller */ |
| 419 | |
| 420 | LRound_ovfl: |
| 421 | /* We are now ready to deal with rounding, but first we must get |
| 422 | the bits properly aligned */ |
| 423 | testb $255,FPU_ovfl_flag /* was the num > denom ? */ |
| 424 | je LRound_precision |
| 425 | |
| 426 | incw EXP(%edi) |
| 427 | |
| 428 | /* shift the mantissa right one bit */ |
| 429 | stc /* Will set the ms bit */ |
| 430 | rcrl FPU_result_2 |
| 431 | rcrl FPU_result_1 |
| 432 | rcrl %eax |
| 433 | |
| 434 | /* Round the result as required */ |
| 435 | LRound_precision: |
| 436 | decw EXP(%edi) /* binary point between 1st & 2nd bits */ |
| 437 | |
| 438 | movl %eax,%edx |
| 439 | movl FPU_result_1,%ebx |
| 440 | movl FPU_result_2,%eax |
| 441 | jmp fpu_reg_round |
| 442 | |
| 443 | |
| 444 | #ifdef PARANOID |
| 445 | /* The logic is wrong if we got here */ |
| 446 | L_bugged: |
| 447 | pushl EX_INTERNAL|0x202 |
| 448 | call EXCEPTION |
| 449 | pop %ebx |
| 450 | jmp L_exit |
| 451 | |
| 452 | L_bugged_1: |
| 453 | pushl EX_INTERNAL|0x203 |
| 454 | call EXCEPTION |
| 455 | pop %ebx |
| 456 | jmp L_exit |
| 457 | |
| 458 | L_bugged_2: |
| 459 | pushl EX_INTERNAL|0x204 |
| 460 | call EXCEPTION |
| 461 | pop %ebx |
| 462 | jmp L_exit |
| 463 | |
| 464 | L_exit: |
| 465 | movl $-1,%eax |
| 466 | popl %ebx |
| 467 | popl %edi |
| 468 | popl %esi |
| 469 | |
| 470 | leave |
| 471 | ret |
| 472 | #endif /* PARANOID */ |
Jiri Slaby | bd6be57 | 2017-08-24 10:06:23 +0200 | [diff] [blame] | 473 | |
Jiri Slaby | 6d685e5 | 2019-10-11 13:51:07 +0200 | [diff] [blame] | 474 | SYM_FUNC_END(FPU_u_div) |