git.oblomov.eu Git - linux-2.6/blob - arch/x86/math-emu/poly_sin.c

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