1 | 2 | decbin.sa 3.3 12/19/90 3 | 4 | Description: Converts normalized packed bcd value pointed to by 5 | register A6 to extended-precision value in FP0. 6 | 7 | Input: Normalized packed bcd value in ETEMP(a6). 8 | 9 | Output: Exact floating-point representation of the packed bcd value. 10 | 11 | Saves and Modifies: D2-D5 12 | 13 | Speed: The program decbin takes ??? cycles to execute. 14 | 15 | Object Size: 16 | 17 | External Reference(s): None. 18 | 19 | Algorithm: 20 | Expected is a normal bcd (i.e. non-exceptional; all inf, zero, 21 | and NaN operands are dispatched without entering this routine) 22 | value in 68881/882 format at location ETEMP(A6). 23 | 24 | A1. Convert the bcd exponent to binary by successive adds and muls. 25 | Set the sign according to SE. Subtract 16 to compensate 26 | for the mantissa which is to be interpreted as 17 integer 27 | digits, rather than 1 integer and 16 fraction digits. 28 | Note: this operation can never overflow. 29 | 30 | A2. Convert the bcd mantissa to binary by successive 31 | adds and muls in FP0. Set the sign according to SM. 32 | The mantissa digits will be converted with the decimal point 33 | assumed following the least-significant digit. 34 | Note: this operation can never overflow. 35 | 36 | A3. Count the number of leading/trailing zeros in the 37 | bcd string. If SE is positive, count the leading zeros; 38 | if negative, count the trailing zeros. Set the adjusted 39 | exponent equal to the exponent from A1 and the zero count 40 | added if SM = 1 and subtracted if SM = 0. Scale the 41 | mantissa the equivalent of forcing in the bcd value: 42 | 43 | SM = 0 a non-zero digit in the integer position 44 | SM = 1 a non-zero digit in Mant0, lsd of the fraction 45 | 46 | this will insure that any value, regardless of its 47 | representation (ex. 0.1E2, 1E1, 10E0, 100E-1), is converted 48 | consistently. 49 | 50 | A4. Calculate the factor 10^exp in FP1 using a table of 51 | 10^(2^n) values. To reduce the error in forming factors 52 | greater than 10^27, a directed rounding scheme is used with 53 | tables rounded to RN, RM, and RP, according to the table 54 | in the comments of the pwrten section. 55 | 56 | A5. Form the final binary number by scaling the mantissa by 57 | the exponent factor. This is done by multiplying the 58 | mantissa in FP0 by the factor in FP1 if the adjusted 59 | exponent sign is positive, and dividing FP0 by FP1 if 60 | it is negative. 61 | 62 | Clean up and return. Check if the final mul or div resulted 63 | in an inex2 exception. If so, set inex1 in the fpsr and 64 | check if the inex1 exception is enabled. If so, set d7 upper 65 | word to $0100. This will signal unimp.sa that an enabled inex1 66 | exception occurred. Unimp will fix the stack. 67 | 68 69 | Copyright (C) Motorola, Inc. 1990 70 | All Rights Reserved 71 | 72 | THIS IS UNPUBLISHED PROPRIETARY SOURCE CODE OF MOTOROLA 73 | The copyright notice above does not evidence any 74 | actual or intended publication of such source code. 75 76 |DECBIN idnt 2,1 | Motorola 040 Floating Point Software Package 77 78 |section 8 79 80 .include "fpsp.h" 81 82 | 83 | PTENRN, PTENRM, and PTENRP are arrays of powers of 10 rounded 84 | to nearest, minus, and plus, respectively. The tables include 85 | 10**{1,2,4,8,16,32,64,128,256,512,1024,2048,4096}. No rounding 86 | is required until the power is greater than 27, however, all 87 | tables include the first 5 for ease of indexing. 88 | 89 |xref PTENRN 90 |xref PTENRM 91 |xref PTENRP 92 93 RTABLE: .byte 0,0,0,0 94 .byte 2,3,2,3 95 .byte 2,3,3,2 96 .byte 3,2,2,3 97 98 .global decbin 99 .global calc_e 100 .global pwrten 101 .global calc_m 102 .global norm 103 .global ap_st_z 104 .global ap_st_n 105 | 106 .set FNIBS,7 107 .set FSTRT,0 108 | 109 .set ESTRT,4 110 .set EDIGITS,2 | 111 | 112 | Constants in single precision 113 FZERO: .long 0x00000000 114 FONE: .long 0x3F800000 115 FTEN: .long 0x41200000 116 117 .set TEN,10 118 119 | 120 decbin: 121 | fmovel #0,FPCR ;clr real fpcr 122 moveml %d2-%d5,-(%a7) 123 | 124 | Calculate exponent: 125 | 1. Copy bcd value in memory for use as a working copy. 126 | 2. Calculate absolute value of exponent in d1 by mul and add. 127 | 3. Correct for exponent sign. 128 | 4. Subtract 16 to compensate for interpreting the mant as all integer digits. 129 | (i.e., all digits assumed left of the decimal point.) 130 | 131 | Register usage: 132 | 133 | calc_e: 134 | (*) d0: temp digit storage 135 | (*) d1: accumulator for binary exponent 136 | (*) d2: digit count 137 | (*) d3: offset pointer 138 | ( ) d4: first word of bcd 139 | ( ) a0: pointer to working bcd value 140 | ( ) a6: pointer to original bcd value 141 | (*) FP_SCR1: working copy of original bcd value 142 | (*) L_SCR1: copy of original exponent word 143 | 144 calc_e: 145 movel #EDIGITS,%d2 |# of nibbles (digits) in fraction part 146 moveql #ESTRT,%d3 |counter to pick up digits 147 leal FP_SCR1(%a6),%a0 |load tmp bcd storage address 148 movel ETEMP(%a6),(%a0) |save input bcd value 149 movel ETEMP_HI(%a6),4(%a0) |save words 2 and 3 150 movel ETEMP_LO(%a6),8(%a0) |and work with these 151 movel (%a0),%d4 |get first word of bcd 152 clrl %d1 |zero d1 for accumulator 153 e_gd: 154 mulul #TEN,%d1 |mul partial product by one digit place 155 bfextu %d4{%d3:#4},%d0 |get the digit and zero extend into d0 156 addl %d0,%d1 |d1 = d1 + d0 157 addqb #4,%d3 |advance d3 to the next digit 158 dbf %d2,e_gd |if we have used all 3 digits, exit loop 159 btst #30,%d4 |get SE 160 beqs e_pos |don't negate if pos 161 negl %d1 |negate before subtracting 162 e_pos: 163 subl #16,%d1 |sub to compensate for shift of mant 164 bges e_save |if still pos, do not neg 165 negl %d1 |now negative, make pos and set SE 166 orl #0x40000000,%d4 |set SE in d4, 167 orl #0x40000000,(%a0) |and in working bcd 168 e_save: 169 movel %d1,L_SCR1(%a6) |save exp in memory 170 | 171 | 172 | Calculate mantissa: 173 | 1. Calculate absolute value of mantissa in fp0 by mul and add. 174 | 2. Correct for mantissa sign. 175 | (i.e., all digits assumed left of the decimal point.) 176 | 177 | Register usage: 178 | 179 | calc_m: 180 | (*) d0: temp digit storage 181 | (*) d1: lword counter 182 | (*) d2: digit count 183 | (*) d3: offset pointer 184 | ( ) d4: words 2 and 3 of bcd 185 | ( ) a0: pointer to working bcd value 186 | ( ) a6: pointer to original bcd value 187 | (*) fp0: mantissa accumulator 188 | ( ) FP_SCR1: working copy of original bcd value 189 | ( ) L_SCR1: copy of original exponent word 190 | 191 calc_m: 192 moveql #1,%d1 |word counter, init to 1 193 fmoves FZERO,%fp0 |accumulator 194 | 195 | 196 | Since the packed number has a long word between the first & second parts, 197 | get the integer digit then skip down & get the rest of the 198 | mantissa. We will unroll the loop once. 199 | 200 bfextu (%a0){#28:#4},%d0 |integer part is ls digit in long word 201 faddb %d0,%fp0 |add digit to sum in fp0 202 | 203 | 204 | Get the rest of the mantissa. 205 | 206 loadlw: 207 movel (%a0,%d1.L*4),%d4 |load mantissa longword into d4 208 moveql #FSTRT,%d3 |counter to pick up digits 209 moveql #FNIBS,%d2 |reset number of digits per a0 ptr 210 md2b: 211 fmuls FTEN,%fp0 |fp0 = fp0 * 10 212 bfextu %d4{%d3:#4},%d0 |get the digit and zero extend 213 faddb %d0,%fp0 |fp0 = fp0 + digit 214 | 215 | 216 | If all the digits (8) in that long word have been converted (d2=0), 217 | then inc d1 (=2) to point to the next long word and reset d3 to 0 218 | to initialize the digit offset, and set d2 to 7 for the digit count; 219 | else continue with this long word. 220 | 221 addqb #4,%d3 |advance d3 to the next digit 222 dbf %d2,md2b |check for last digit in this lw 223 nextlw: 224 addql #1,%d1 |inc lw pointer in mantissa 225 cmpl #2,%d1 |test for last lw 226 ble loadlw |if not, get last one 227 228 | 229 | Check the sign of the mant and make the value in fp0 the same sign. 230 | 231 m_sign: 232 btst #31,(%a0) |test sign of the mantissa 233 beqs ap_st_z |if clear, go to append/strip zeros 234 fnegx %fp0 |if set, negate fp0 235 236 | 237 | Append/strip zeros: 238 | 239 | For adjusted exponents which have an absolute value greater than 27*, 240 | this routine calculates the amount needed to normalize the mantissa 241 | for the adjusted exponent. That number is subtracted from the exp 242 | if the exp was positive, and added if it was negative. The purpose 243 | of this is to reduce the value of the exponent and the possibility 244 | of error in calculation of pwrten. 245 | 246 | 1. Branch on the sign of the adjusted exponent. 247 | 2p.(positive exp) 248 | 2. Check M16 and the digits in lwords 2 and 3 in descending order. 249 | 3. Add one for each zero encountered until a non-zero digit. 250 | 4. Subtract the count from the exp. 251 | 5. Check if the exp has crossed zero in #3 above; make the exp abs 252 | and set SE. 253 | 6. Multiply the mantissa by 10**count. 254 | 2n.(negative exp) 255 | 2. Check the digits in lwords 3 and 2 in descending order. 256 | 3. Add one for each zero encountered until a non-zero digit. 257 | 4. Add the count to the exp. 258 | 5. Check if the exp has crossed zero in #3 above; clear SE. 259 | 6. Divide the mantissa by 10**count. 260 | 261 | *Why 27? If the adjusted exponent is within -28 < expA < 28, than 262 | any adjustment due to append/strip zeros will drive the resultant 263 | exponent towards zero. Since all pwrten constants with a power 264 | of 27 or less are exact, there is no need to use this routine to 265 | attempt to lessen the resultant exponent. 266 | 267 | Register usage: 268 | 269 | ap_st_z: 270 | (*) d0: temp digit storage 271 | (*) d1: zero count 272 | (*) d2: digit count 273 | (*) d3: offset pointer 274 | ( ) d4: first word of bcd 275 | (*) d5: lword counter 276 | ( ) a0: pointer to working bcd value 277 | ( ) FP_SCR1: working copy of original bcd value 278 | ( ) L_SCR1: copy of original exponent word 279 | 280 | 281 | First check the absolute value of the exponent to see if this 282 | routine is necessary. If so, then check the sign of the exponent 283 | and do append (+) or strip (-) zeros accordingly. 284 | This section handles a positive adjusted exponent. 285 | 286 ap_st_z: 287 movel L_SCR1(%a6),%d1 |load expA for range test 288 cmpl #27,%d1 |test is with 27 289 ble pwrten |if abs(expA) <28, skip ap/st zeros 290 btst #30,(%a0) |check sign of exp 291 bnes ap_st_n |if neg, go to neg side 292 clrl %d1 |zero count reg 293 movel (%a0),%d4 |load lword 1 to d4 294 bfextu %d4{#28:#4},%d0 |get M16 in d0 295 bnes ap_p_fx |if M16 is non-zero, go fix exp 296 addql #1,%d1 |inc zero count 297 moveql #1,%d5 |init lword counter 298 movel (%a0,%d5.L*4),%d4 |get lword 2 to d4 299 bnes ap_p_cl |if lw 2 is zero, skip it 300 addql #8,%d1 |and inc count by 8 301 addql #1,%d5 |inc lword counter 302 movel (%a0,%d5.L*4),%d4 |get lword 3 to d4 303 ap_p_cl: 304 clrl %d3 |init offset reg 305 moveql #7,%d2 |init digit counter 306 ap_p_gd: 307 bfextu %d4{%d3:#4},%d0 |get digit 308 bnes ap_p_fx |if non-zero, go to fix exp 309 addql #4,%d3 |point to next digit 310 addql #1,%d1 |inc digit counter 311 dbf %d2,ap_p_gd |get next digit 312 ap_p_fx: 313 movel %d1,%d0 |copy counter to d2 314 movel L_SCR1(%a6),%d1 |get adjusted exp from memory 315 subl %d0,%d1 |subtract count from exp 316 bges ap_p_fm |if still pos, go to pwrten 317 negl %d1 |now its neg; get abs 318 movel (%a0),%d4 |load lword 1 to d4 319 orl #0x40000000,%d4 | and set SE in d4 320 orl #0x40000000,(%a0) | and in memory 321 | 322 | Calculate the mantissa multiplier to compensate for the striping of 323 | zeros from the mantissa. 324 | 325 ap_p_fm: 326 movel #PTENRN,%a1 |get address of power-of-ten table 327 clrl %d3 |init table index 328 fmoves FONE,%fp1 |init fp1 to 1 329 moveql #3,%d2 |init d2 to count bits in counter 330 ap_p_el: 331 asrl #1,%d0 |shift lsb into carry 332 bccs ap_p_en |if 1, mul fp1 by pwrten factor 333 fmulx (%a1,%d3),%fp1 |mul by 10**(d3_bit_no) 334 ap_p_en: 335 addl #12,%d3 |inc d3 to next rtable entry 336 tstl %d0 |check if d0 is zero 337 bnes ap_p_el |if not, get next bit 338 fmulx %fp1,%fp0 |mul mantissa by 10**(no_bits_shifted) 339 bras pwrten |go calc pwrten 340 | 341 | This section handles a negative adjusted exponent. 342 | 343 ap_st_n: 344 clrl %d1 |clr counter 345 moveql #2,%d5 |set up d5 to point to lword 3 346 movel (%a0,%d5.L*4),%d4 |get lword 3 347 bnes ap_n_cl |if not zero, check digits 348 subl #1,%d5 |dec d5 to point to lword 2 349 addql #8,%d1 |inc counter by 8 350 movel (%a0,%d5.L*4),%d4 |get lword 2 351 ap_n_cl: 352 movel #28,%d3 |point to last digit 353 moveql #7,%d2 |init digit counter 354 ap_n_gd: 355 bfextu %d4{%d3:#4},%d0 |get digit 356 bnes ap_n_fx |if non-zero, go to exp fix 357 subql #4,%d3 |point to previous digit 358 addql #1,%d1 |inc digit counter 359 dbf %d2,ap_n_gd |get next digit 360 ap_n_fx: 361 movel %d1,%d0 |copy counter to d0 362 movel L_SCR1(%a6),%d1 |get adjusted exp from memory 363 subl %d0,%d1 |subtract count from exp 364 bgts ap_n_fm |if still pos, go fix mantissa 365 negl %d1 |take abs of exp and clr SE 366 movel (%a0),%d4 |load lword 1 to d4 367 andl #0xbfffffff,%d4 | and clr SE in d4 368 andl #0xbfffffff,(%a0) | and in memory 369 | 370 | Calculate the mantissa multiplier to compensate for the appending of 371 | zeros to the mantissa. 372 | 373 ap_n_fm: 374 movel #PTENRN,%a1 |get address of power-of-ten table 375 clrl %d3 |init table index 376 fmoves FONE,%fp1 |init fp1 to 1 377 moveql #3,%d2 |init d2 to count bits in counter 378 ap_n_el: 379 asrl #1,%d0 |shift lsb into carry 380 bccs ap_n_en |if 1, mul fp1 by pwrten factor 381 fmulx (%a1,%d3),%fp1 |mul by 10**(d3_bit_no) 382 ap_n_en: 383 addl #12,%d3 |inc d3 to next rtable entry 384 tstl %d0 |check if d0 is zero 385 bnes ap_n_el |if not, get next bit 386 fdivx %fp1,%fp0 |div mantissa by 10**(no_bits_shifted) 387 | 388 | 389 | Calculate power-of-ten factor from adjusted and shifted exponent. 390 | 391 | Register usage: 392 | 393 | pwrten: 394 | (*) d0: temp 395 | ( ) d1: exponent 396 | (*) d2: {FPCR[6:5],SM,SE} as index in RTABLE; temp 397 | (*) d3: FPCR work copy 398 | ( ) d4: first word of bcd 399 | (*) a1: RTABLE pointer 400 | calc_p: 401 | (*) d0: temp 402 | ( ) d1: exponent 403 | (*) d3: PWRTxx table index 404 | ( ) a0: pointer to working copy of bcd 405 | (*) a1: PWRTxx pointer 406 | (*) fp1: power-of-ten accumulator 407 | 408 | Pwrten calculates the exponent factor in the selected rounding mode 409 | according to the following table: 410 | 411 | Sign of Mant Sign of Exp Rounding Mode PWRTEN Rounding Mode 412 | 413 | ANY ANY RN RN 414 | 415 | + + RP RP 416 | - + RP RM 417 | + - RP RM 418 | - - RP RP 419 | 420 | + + RM RM 421 | - + RM RP 422 | + - RM RP 423 | - - RM RM 424 | 425 | + + RZ RM 426 | - + RZ RM 427 | + - RZ RP 428 | - - RZ RP 429 | 430 | 431 pwrten: 432 movel USER_FPCR(%a6),%d3 |get user's FPCR 433 bfextu %d3{#26:#2},%d2 |isolate rounding mode bits 434 movel (%a0),%d4 |reload 1st bcd word to d4 435 asll #2,%d2 |format d2 to be 436 bfextu %d4{#0:#2},%d0 | {FPCR[6],FPCR[5],SM,SE} 437 addl %d0,%d2 |in d2 as index into RTABLE 438 leal RTABLE,%a1 |load rtable base 439 moveb (%a1,%d2),%d0 |load new rounding bits from table 440 clrl %d3 |clear d3 to force no exc and extended 441 bfins %d0,%d3{#26:#2} |stuff new rounding bits in FPCR 442 fmovel %d3,%FPCR |write new FPCR 443 asrl #1,%d0 |write correct PTENxx table 444 bccs not_rp |to a1 445 leal PTENRP,%a1 |it is RP 446 bras calc_p |go to init section 447 not_rp: 448 asrl #1,%d0 |keep checking 449 bccs not_rm 450 leal PTENRM,%a1 |it is RM 451 bras calc_p |go to init section 452 not_rm: 453 leal PTENRN,%a1 |it is RN 454 calc_p: 455 movel %d1,%d0 |copy exp to d0;use d0 456 bpls no_neg |if exp is negative, 457 negl %d0 |invert it 458 orl #0x40000000,(%a0) |and set SE bit 459 no_neg: 460 clrl %d3 |table index 461 fmoves FONE,%fp1 |init fp1 to 1 462 e_loop: 463 asrl #1,%d0 |shift next bit into carry 464 bccs e_next |if zero, skip the mul 465 fmulx (%a1,%d3),%fp1 |mul by 10**(d3_bit_no) 466 e_next: 467 addl #12,%d3 |inc d3 to next rtable entry 468 tstl %d0 |check if d0 is zero 469 bnes e_loop |not zero, continue shifting 470 | 471 | 472 | Check the sign of the adjusted exp and make the value in fp0 the 473 | same sign. If the exp was pos then multiply fp1*fp0; 474 | else divide fp0/fp1. 475 | 476 | Register Usage: 477 | norm: 478 | ( ) a0: pointer to working bcd value 479 | (*) fp0: mantissa accumulator 480 | ( ) fp1: scaling factor - 10**(abs(exp)) 481 | 482 norm: 483 btst #30,(%a0) |test the sign of the exponent 484 beqs mul |if clear, go to multiply 485 div: 486 fdivx %fp1,%fp0 |exp is negative, so divide mant by exp 487 bras end_dec 488 mul: 489 fmulx %fp1,%fp0 |exp is positive, so multiply by exp 490 | 491 | 492 | Clean up and return with result in fp0. 493 | 494 | If the final mul/div in decbin incurred an inex exception, 495 | it will be inex2, but will be reported as inex1 by get_op. 496 | 497 end_dec: 498 fmovel %FPSR,%d0 |get status register 499 bclrl #inex2_bit+8,%d0 |test for inex2 and clear it 500 fmovel %d0,%FPSR |return status reg w/o inex2 501 beqs no_exc |skip this if no exc 502 orl #inx1a_mask,USER_FPSR(%a6) |set inex1/ainex 503 no_exc: 504 moveml (%a7)+,%d2-%d5 505 rts 506 |end