arch/x86/math-emu/poly_sin.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*---------------------------------------------------------------------------+
   3  |  poly_sin.c                                                               |
   4  |                                                                           |
   5  |  Computation of an approximation of the sin function and the cosine       |
   6  |  function by a polynomial.                                                |
   7  |                                                                           |
   8  | Copyright (C) 1992,1993,1994,1997,1999                                    |
   9  |                  W. Metzenthen, 22 Parker St, Ormond, Vic 3163, Australia |
  10  |                  E-mail   billm@melbpc.org.au                             |
  11  |                                                                           |
  12  |                                                                           |
  13  +---------------------------------------------------------------------------*/
  14
  15 #include "exception.h"
  16 #include "reg_constant.h"
  17 #include "fpu_emu.h"
  18 #include "fpu_system.h"
  19 #include "control_w.h"
  20 #include "poly.h"
  21
  22 #define N_COEFF_P       4
  23 #define N_COEFF_N       4
  24
  25 static const unsigned long long pos_terms_l[N_COEFF_P] = {
  26         0xaaaaaaaaaaaaaaabLL,
  27         0x00d00d00d00cf906LL,
  28         0x000006b99159a8bbLL,
  29         0x000000000d7392e6LL
  30 };
  31
  32 static const unsigned long long neg_terms_l[N_COEFF_N] = {
  33         0x2222222222222167LL,
  34         0x0002e3bc74aab624LL,
  35         0x0000000b09229062LL,
  36         0x00000000000c7973LL
  37 };
  38
  39 #define N_COEFF_PH      4
  40 #define N_COEFF_NH      4
  41 static const unsigned long long pos_terms_h[N_COEFF_PH] = {
  42         0x0000000000000000LL,
  43         0x05b05b05b05b0406LL,
  44         0x000049f93edd91a9LL,
  45         0x00000000c9c9ed62LL
  46 };
  47
  48 static const unsigned long long neg_terms_h[N_COEFF_NH] = {
  49         0xaaaaaaaaaaaaaa98LL,
  50         0x001a01a01a019064LL,
  51         0x0000008f76c68a77LL,
  52         0x0000000000d58f5eLL
  53 };
  54
  55 /*--- poly_sine() -----------------------------------------------------------+
  56  |                                                                           |
  57  +---------------------------------------------------------------------------*/
  58 void poly_sine(FPU_REG *st0_ptr)
  59 {
  60         int exponent, echange;
  61         Xsig accumulator, argSqrd, argTo4;
  62         unsigned long fix_up, adj;
  63         unsigned long long fixed_arg;
  64         FPU_REG result;
  65
  66         exponent = exponent(st0_ptr);
  67
  68         accumulator.lsw = accumulator.midw = accumulator.msw = 0;
  69
  70         /* Split into two ranges, for arguments below and above 1.0 */
  71         /* The boundary between upper and lower is approx 0.88309101259 */
  72         if ((exponent < -1)
  73             || ((exponent == -1) && (st0_ptr->sigh <= 0xe21240aa))) {
  74                 /* The argument is <= 0.88309101259 */
  75
  76                 argSqrd.msw = st0_ptr->sigh;
  77                 argSqrd.midw = st0_ptr->sigl;
  78                 argSqrd.lsw = 0;
  79                 mul64_Xsig(&argSqrd, &significand(st0_ptr));
  80                 shr_Xsig(&argSqrd, 2 * (-1 - exponent));
  81                 argTo4.msw = argSqrd.msw;
  82                 argTo4.midw = argSqrd.midw;
  83                 argTo4.lsw = argSqrd.lsw;
  84                 mul_Xsig_Xsig(&argTo4, &argTo4);
  85
  86                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), neg_terms_l,
  87                                 N_COEFF_N - 1);
  88                 mul_Xsig_Xsig(&accumulator, &argSqrd);
  89                 negate_Xsig(&accumulator);
  90
  91                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), pos_terms_l,
  92                                 N_COEFF_P - 1);
  93
  94                 shr_Xsig(&accumulator, 2);      /* Divide by four */
  95                 accumulator.msw |= 0x80000000;  /* Add 1.0 */
  96
  97                 mul64_Xsig(&accumulator, &significand(st0_ptr));
  98                 mul64_Xsig(&accumulator, &significand(st0_ptr));
  99                 mul64_Xsig(&accumulator, &significand(st0_ptr));
 100
 101                 /* Divide by four, FPU_REG compatible, etc */
 102                 exponent = 3 * exponent;
 103
 104                 /* The minimum exponent difference is 3 */
 105                 shr_Xsig(&accumulator, exponent(st0_ptr) - exponent);
 106
 107                 negate_Xsig(&accumulator);
 108                 XSIG_LL(accumulator) += significand(st0_ptr);
 109
 110                 echange = round_Xsig(&accumulator);
 111
 112                 setexponentpos(&result, exponent(st0_ptr) + echange);
 113         } else {
 114                 /* The argument is > 0.88309101259 */
 115                 /* We use sin(st(0)) = cos(pi/2-st(0)) */
 116
 117                 fixed_arg = significand(st0_ptr);
 118
 119                 if (exponent == 0) {
 120                         /* The argument is >= 1.0 */
 121
 122                         /* Put the binary point at the left. */
 123                         fixed_arg <<= 1;
 124                 }
 125                 /* pi/2 in hex is: 1.921fb54442d18469 898CC51701B839A2 52049C1 */
 126                 fixed_arg = 0x921fb54442d18469LL - fixed_arg;
 127                 /* There is a special case which arises due to rounding, to fix here. */
 128                 if (fixed_arg == 0xffffffffffffffffLL)
 129                         fixed_arg = 0;
 130
 131                 XSIG_LL(argSqrd) = fixed_arg;
 132                 argSqrd.lsw = 0;
 133                 mul64_Xsig(&argSqrd, &fixed_arg);
 134
 135                 XSIG_LL(argTo4) = XSIG_LL(argSqrd);
 136                 argTo4.lsw = argSqrd.lsw;
 137                 mul_Xsig_Xsig(&argTo4, &argTo4);
 138
 139                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), neg_terms_h,
 140                                 N_COEFF_NH - 1);
 141                 mul_Xsig_Xsig(&accumulator, &argSqrd);
 142                 negate_Xsig(&accumulator);
 143
 144                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), pos_terms_h,
 145                                 N_COEFF_PH - 1);
 146                 negate_Xsig(&accumulator);
 147
 148                 mul64_Xsig(&accumulator, &fixed_arg);
 149                 mul64_Xsig(&accumulator, &fixed_arg);
 150
 151                 shr_Xsig(&accumulator, 3);
 152                 negate_Xsig(&accumulator);
 153
 154                 add_Xsig_Xsig(&accumulator, &argSqrd);
 155
 156                 shr_Xsig(&accumulator, 1);
 157
 158                 accumulator.lsw |= 1;   /* A zero accumulator here would cause problems */
 159                 negate_Xsig(&accumulator);
 160
 161                 /* The basic computation is complete. Now fix the answer to
 162                    compensate for the error due to the approximation used for
 163                    pi/2
 164                  */
 165
 166                 /* This has an exponent of -65 */
 167                 fix_up = 0x898cc517;
 168                 /* The fix-up needs to be improved for larger args */
 169                 if (argSqrd.msw & 0xffc00000) {
 170                         /* Get about 32 bit precision in these: */
 171                         fix_up -= mul_32_32(0x898cc517, argSqrd.msw) / 6;
 172                 }
 173                 fix_up = mul_32_32(fix_up, LL_MSW(fixed_arg));
 174
 175                 adj = accumulator.lsw;  /* temp save */
 176                 accumulator.lsw -= fix_up;
 177                 if (accumulator.lsw > adj)
 178                         XSIG_LL(accumulator)--;
 179
 180                 echange = round_Xsig(&accumulator);
 181
 182                 setexponentpos(&result, echange - 1);
 183         }
 184
 185         significand(&result) = XSIG_LL(accumulator);
 186         setsign(&result, getsign(st0_ptr));
 187         FPU_copy_to_reg0(&result, TAG_Valid);
 188
 189 #ifdef PARANOID
 190         if ((exponent(&result) >= 0)
 191             && (significand(&result) > 0x8000000000000000LL)) {
 192                 EXCEPTION(EX_INTERNAL | 0x150);
 193         }
 194 #endif /* PARANOID */
 195
 196 }
 197
 198 /*--- poly_cos() ------------------------------------------------------------+
 199  |                                                                           |
 200  +---------------------------------------------------------------------------*/
 201 void poly_cos(FPU_REG *st0_ptr)
 202 {
 203         FPU_REG result;
 204         long int exponent, exp2, echange;
 205         Xsig accumulator, argSqrd, fix_up, argTo4;
 206         unsigned long long fixed_arg;
 207
 208 #ifdef PARANOID
 209         if ((exponent(st0_ptr) > 0)
 210             || ((exponent(st0_ptr) == 0)
 211                 && (significand(st0_ptr) > 0xc90fdaa22168c234LL))) {
 212                 EXCEPTION(EX_Invalid);
 213                 FPU_copy_to_reg0(&CONST_QNaN, TAG_Special);
 214                 return;
 215         }
 216 #endif /* PARANOID */
 217
 218         exponent = exponent(st0_ptr);
 219
 220         accumulator.lsw = accumulator.midw = accumulator.msw = 0;
 221
 222         if ((exponent < -1)
 223             || ((exponent == -1) && (st0_ptr->sigh <= 0xb00d6f54))) {
 224                 /* arg is < 0.687705 */
 225
 226                 argSqrd.msw = st0_ptr->sigh;
 227                 argSqrd.midw = st0_ptr->sigl;
 228                 argSqrd.lsw = 0;
 229                 mul64_Xsig(&argSqrd, &significand(st0_ptr));
 230
 231                 if (exponent < -1) {
 232                         /* shift the argument right by the required places */
 233                         shr_Xsig(&argSqrd, 2 * (-1 - exponent));
 234                 }
 235
 236                 argTo4.msw = argSqrd.msw;
 237                 argTo4.midw = argSqrd.midw;
 238                 argTo4.lsw = argSqrd.lsw;
 239                 mul_Xsig_Xsig(&argTo4, &argTo4);
 240
 241                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), neg_terms_h,
 242                                 N_COEFF_NH - 1);
 243                 mul_Xsig_Xsig(&accumulator, &argSqrd);
 244                 negate_Xsig(&accumulator);
 245
 246                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), pos_terms_h,
 247                                 N_COEFF_PH - 1);
 248                 negate_Xsig(&accumulator);
 249
 250                 mul64_Xsig(&accumulator, &significand(st0_ptr));
 251                 mul64_Xsig(&accumulator, &significand(st0_ptr));
 252                 shr_Xsig(&accumulator, -2 * (1 + exponent));
 253
 254                 shr_Xsig(&accumulator, 3);
 255                 negate_Xsig(&accumulator);
 256
 257                 add_Xsig_Xsig(&accumulator, &argSqrd);
 258
 259                 shr_Xsig(&accumulator, 1);
 260
 261                 /* It doesn't matter if accumulator is all zero here, the
 262                    following code will work ok */
 263                 negate_Xsig(&accumulator);
 264
 265                 if (accumulator.lsw & 0x80000000)
 266                         XSIG_LL(accumulator)++;
 267                 if (accumulator.msw == 0) {
 268                         /* The result is 1.0 */
 269                         FPU_copy_to_reg0(&CONST_1, TAG_Valid);
 270                         return;
 271                 } else {
 272                         significand(&result) = XSIG_LL(accumulator);
 273
 274                         /* will be a valid positive nr with expon = -1 */
 275                         setexponentpos(&result, -1);
 276                 }
 277         } else {
 278                 fixed_arg = significand(st0_ptr);
 279
 280                 if (exponent == 0) {
 281                         /* The argument is >= 1.0 */
 282
 283                         /* Put the binary point at the left. */
 284                         fixed_arg <<= 1;
 285                 }
 286                 /* pi/2 in hex is: 1.921fb54442d18469 898CC51701B839A2 52049C1 */
 287                 fixed_arg = 0x921fb54442d18469LL - fixed_arg;
 288                 /* There is a special case which arises due to rounding, to fix here. */
 289                 if (fixed_arg == 0xffffffffffffffffLL)
 290                         fixed_arg = 0;
 291
 292                 exponent = -1;
 293                 exp2 = -1;
 294
 295                 /* A shift is needed here only for a narrow range of arguments,
 296                    i.e. for fixed_arg approx 2^-32, but we pick up more... */
 297                 if (!(LL_MSW(fixed_arg) & 0xffff0000)) {
 298                         fixed_arg <<= 16;
 299                         exponent -= 16;
 300                         exp2 -= 16;
 301                 }
 302
 303                 XSIG_LL(argSqrd) = fixed_arg;
 304                 argSqrd.lsw = 0;
 305                 mul64_Xsig(&argSqrd, &fixed_arg);
 306
 307                 if (exponent < -1) {
 308                         /* shift the argument right by the required places */
 309                         shr_Xsig(&argSqrd, 2 * (-1 - exponent));
 310                 }
 311
 312                 argTo4.msw = argSqrd.msw;
 313                 argTo4.midw = argSqrd.midw;
 314                 argTo4.lsw = argSqrd.lsw;
 315                 mul_Xsig_Xsig(&argTo4, &argTo4);
 316
 317                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), neg_terms_l,
 318                                 N_COEFF_N - 1);
 319                 mul_Xsig_Xsig(&accumulator, &argSqrd);
 320                 negate_Xsig(&accumulator);
 321
 322                 polynomial_Xsig(&accumulator, &XSIG_LL(argTo4), pos_terms_l,
 323                                 N_COEFF_P - 1);
 324
 325                 shr_Xsig(&accumulator, 2);      /* Divide by four */
 326                 accumulator.msw |= 0x80000000;  /* Add 1.0 */
 327
 328                 mul64_Xsig(&accumulator, &fixed_arg);
 329                 mul64_Xsig(&accumulator, &fixed_arg);
 330                 mul64_Xsig(&accumulator, &fixed_arg);
 331
 332                 /* Divide by four, FPU_REG compatible, etc */
 333                 exponent = 3 * exponent;
 334
 335                 /* The minimum exponent difference is 3 */
 336                 shr_Xsig(&accumulator, exp2 - exponent);
 337
 338                 negate_Xsig(&accumulator);
 339                 XSIG_LL(accumulator) += fixed_arg;
 340
 341                 /* The basic computation is complete. Now fix the answer to
 342                    compensate for the error due to the approximation used for
 343                    pi/2
 344                  */
 345
 346                 /* This has an exponent of -65 */
 347                 XSIG_LL(fix_up) = 0x898cc51701b839a2ll;
 348                 fix_up.lsw = 0;
 349
 350                 /* The fix-up needs to be improved for larger args */
 351                 if (argSqrd.msw & 0xffc00000) {
 352                         /* Get about 32 bit precision in these: */
 353                         fix_up.msw -= mul_32_32(0x898cc517, argSqrd.msw) / 2;
 354                         fix_up.msw += mul_32_32(0x898cc517, argTo4.msw) / 24;
 355                 }
 356
 357                 exp2 += norm_Xsig(&accumulator);
 358                 shr_Xsig(&accumulator, 1);      /* Prevent overflow */
 359                 exp2++;
 360                 shr_Xsig(&fix_up, 65 + exp2);
 361
 362                 add_Xsig_Xsig(&accumulator, &fix_up);
 363
 364                 echange = round_Xsig(&accumulator);
 365
 366                 setexponentpos(&result, exp2 + echange);
 367                 significand(&result) = XSIG_LL(accumulator);
 368         }
 369
 370         FPU_copy_to_reg0(&result, TAG_Valid);
 371
 372 #ifdef PARANOID
 373         if ((exponent(&result) >= 0)
 374             && (significand(&result) > 0x8000000000000000LL)) {
 375                 EXCEPTION(EX_INTERNAL | 0x151);
 376         }
 377 #endif /* PARANOID */
 378
 379 }