git.oblomov.eu Git - linux-2.6/blob - arch/arm/vfp/vfpdouble.c

   1 /*
   2  *  linux/arch/arm/vfp/vfpdouble.c
   3  *
   4  * This code is derived in part from John R. Housers softfloat library, which
   5  * carries the following notice:
   6  *
   7  * ===========================================================================
   8  * This C source file is part of the SoftFloat IEC/IEEE Floating-point
   9  * Arithmetic Package, Release 2.
  10  *
  11  * Written by John R. Hauser.  This work was made possible in part by the
  12  * International Computer Science Institute, located at Suite 600, 1947 Center
  13  * Street, Berkeley, California 94704.  Funding was partially provided by the
  14  * National Science Foundation under grant MIP-9311980.  The original version
  15  * of this code was written as part of a project to build a fixed-point vector
  16  * processor in collaboration with the University of California at Berkeley,
  17  * overseen by Profs. Nelson Morgan and John Wawrzynek.  More information
  18  * is available through the web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
  19  * arithmetic/softfloat.html'.
  20  *
  21  * THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE.  Although reasonable effort
  22  * has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
  23  * TIMES RESULT IN INCORRECT BEHAVIOR.  USE OF THIS SOFTWARE IS RESTRICTED TO
  24  * PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
  25  * AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
  26  *
  27  * Derivative works are acceptable, even for commercial purposes, so long as
  28  * (1) they include prominent notice that the work is derivative, and (2) they
  29  * include prominent notice akin to these three paragraphs for those parts of
  30  * this code that are retained.
  31  * ===========================================================================
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/bitops.h>
  35
  36 #include <asm/div64.h>
  37 #include <asm/ptrace.h>
  38 #include <asm/vfp.h>
  39
  40 #include "vfpinstr.h"
  41 #include "vfp.h"
  42
  43 static struct vfp_double vfp_double_default_qnan = {
  44         .exponent       = 2047,
  45         .sign           = 0,
  46         .significand    = VFP_DOUBLE_SIGNIFICAND_QNAN,
  47 };
  48
  49 static void vfp_double_dump(const char *str, struct vfp_double *d)
  50 {
  51         pr_debug("VFP: %s: sign=%d exponent=%d significand=%016llx\n",
  52                  str, d->sign != 0, d->exponent, d->significand);
  53 }
  54
  55 static void vfp_double_normalise_denormal(struct vfp_double *vd)
  56 {
  57         int bits = 31 - fls(vd->significand >> 32);
  58         if (bits == 31)
  59                 bits = 62 - fls(vd->significand);
  60
  61         vfp_double_dump("normalise_denormal: in", vd);
  62
  63         if (bits) {
  64                 vd->exponent -= bits - 1;
  65                 vd->significand <<= bits;
  66         }
  67
  68         vfp_double_dump("normalise_denormal: out", vd);
  69 }
  70
  71 u32 vfp_double_normaliseround(int dd, struct vfp_double *vd, u32 fpscr, u32 exceptions, const char *func)
  72 {
  73         u64 significand, incr;
  74         int exponent, shift, underflow;
  75         u32 rmode;
  76
  77         vfp_double_dump("pack: in", vd);
  78
  79         /*
  80          * Infinities and NaNs are a special case.
  81          */
  82         if (vd->exponent == 2047 && (vd->significand == 0 || exceptions))
  83                 goto pack;
  84
  85         /*
  86          * Special-case zero.
  87          */
  88         if (vd->significand == 0) {
  89                 vd->exponent = 0;
  90                 goto pack;
  91         }
  92
  93         exponent = vd->exponent;
  94         significand = vd->significand;
  95
  96         shift = 32 - fls(significand >> 32);
  97         if (shift == 32)
  98                 shift = 64 - fls(significand);
  99         if (shift) {
 100                 exponent -= shift;
 101                 significand <<= shift;
 102         }
 103
 104 #ifdef DEBUG
 105         vd->exponent = exponent;
 106         vd->significand = significand;
 107         vfp_double_dump("pack: normalised", vd);
 108 #endif
 109
 110         /*
 111          * Tiny number?
 112          */
 113         underflow = exponent < 0;
 114         if (underflow) {
 115                 significand = vfp_shiftright64jamming(significand, -exponent);
 116                 exponent = 0;
 117 #ifdef DEBUG
 118                 vd->exponent = exponent;
 119                 vd->significand = significand;
 120                 vfp_double_dump("pack: tiny number", vd);
 121 #endif
 122                 if (!(significand & ((1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1)))
 123                         underflow = 0;
 124         }
 125
 126         /*
 127          * Select rounding increment.
 128          */
 129         incr = 0;
 130         rmode = fpscr & FPSCR_RMODE_MASK;
 131
 132         if (rmode == FPSCR_ROUND_NEAREST) {
 133                 incr = 1ULL << VFP_DOUBLE_LOW_BITS;
 134                 if ((significand & (1ULL << (VFP_DOUBLE_LOW_BITS + 1))) == 0)
 135                         incr -= 1;
 136         } else if (rmode == FPSCR_ROUND_TOZERO) {
 137                 incr = 0;
 138         } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vd->sign != 0))
 139                 incr = (1ULL << (VFP_DOUBLE_LOW_BITS + 1)) - 1;
 140
 141         pr_debug("VFP: rounding increment = 0x%08llx\n", incr);
 142
 143         /*
 144          * Is our rounding going to overflow?
 145          */
 146         if ((significand + incr) < significand) {
 147                 exponent += 1;
 148                 significand = (significand >> 1) | (significand & 1);
 149                 incr >>= 1;
 150 #ifdef DEBUG
 151                 vd->exponent = exponent;
 152                 vd->significand = significand;
 153                 vfp_double_dump("pack: overflow", vd);
 154 #endif
 155         }
 156
 157         /*
 158          * If any of the low bits (which will be shifted out of the
 159          * number) are non-zero, the result is inexact.
 160          */
 161         if (significand & ((1 << (VFP_DOUBLE_LOW_BITS + 1)) - 1))
 162                 exceptions |= FPSCR_IXC;
 163
 164         /*
 165          * Do our rounding.
 166          */
 167         significand += incr;
 168
 169         /*
 170          * Infinity?
 171          */
 172         if (exponent >= 2046) {
 173                 exceptions |= FPSCR_OFC | FPSCR_IXC;
 174                 if (incr == 0) {
 175                         vd->exponent = 2045;
 176                         vd->significand = 0x7fffffffffffffffULL;
 177                 } else {
 178                         vd->exponent = 2047;            /* infinity */
 179                         vd->significand = 0;
 180                 }
 181         } else {
 182                 if (significand >> (VFP_DOUBLE_LOW_BITS + 1) == 0)
 183                         exponent = 0;
 184                 if (exponent || significand > 0x8000000000000000ULL)
 185                         underflow = 0;
 186                 if (underflow)
 187                         exceptions |= FPSCR_UFC;
 188                 vd->exponent = exponent;
 189                 vd->significand = significand >> 1;
 190         }
 191
 192  pack:
 193         vfp_double_dump("pack: final", vd);
 194         {
 195                 s64 d = vfp_double_pack(vd);
 196                 pr_debug("VFP: %s: d(d%d)=%016llx exceptions=%08x\n", func,
 197                          dd, d, exceptions);
 198                 vfp_put_double(dd, d);
 199         }
 200         return exceptions & ~VFP_NAN_FLAG;
 201 }
 202
 203 /*
 204  * Propagate the NaN, setting exceptions if it is signalling.
 205  * 'n' is always a NaN.  'm' may be a number, NaN or infinity.
 206  */
 207 static u32
 208 vfp_propagate_nan(struct vfp_double *vdd, struct vfp_double *vdn,
 209                   struct vfp_double *vdm, u32 fpscr)
 210 {
 211         struct vfp_double *nan;
 212         int tn, tm = 0;
 213
 214         tn = vfp_double_type(vdn);
 215
 216         if (vdm)
 217                 tm = vfp_double_type(vdm);
 218
 219         if (fpscr & FPSCR_DEFAULT_NAN)
 220                 /*
 221                  * Default NaN mode - always returns a quiet NaN
 222                  */
 223                 nan = &vfp_double_default_qnan;
 224         else {
 225                 /*
 226                  * Contemporary mode - select the first signalling
 227                  * NAN, or if neither are signalling, the first
 228                  * quiet NAN.
 229                  */
 230                 if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
 231                         nan = vdn;
 232                 else
 233                         nan = vdm;
 234                 /*
 235                  * Make the NaN quiet.
 236                  */
 237                 nan->significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
 238         }
 239
 240         *vdd = *nan;
 241
 242         /*
 243          * If one was a signalling NAN, raise invalid operation.
 244          */
 245         return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
 246 }
 247
 248 /*
 249  * Extended operations
 250  */
 251 static u32 vfp_double_fabs(int dd, int unused, int dm, u32 fpscr)
 252 {
 253         vfp_put_double(dd, vfp_double_packed_abs(vfp_get_double(dm)));
 254         return 0;
 255 }
 256
 257 static u32 vfp_double_fcpy(int dd, int unused, int dm, u32 fpscr)
 258 {
 259         vfp_put_double(dd, vfp_get_double(dm));
 260         return 0;
 261 }
 262
 263 static u32 vfp_double_fneg(int dd, int unused, int dm, u32 fpscr)
 264 {
 265         vfp_put_double(dd, vfp_double_packed_negate(vfp_get_double(dm)));
 266         return 0;
 267 }
 268
 269 static u32 vfp_double_fsqrt(int dd, int unused, int dm, u32 fpscr)
 270 {
 271         struct vfp_double vdm, vdd;
 272         int ret, tm;
 273
 274         vfp_double_unpack(&vdm, vfp_get_double(dm));
 275         tm = vfp_double_type(&vdm);
 276         if (tm & (VFP_NAN|VFP_INFINITY)) {
 277                 struct vfp_double *vdp = &vdd;
 278
 279                 if (tm & VFP_NAN)
 280                         ret = vfp_propagate_nan(vdp, &vdm, NULL, fpscr);
 281                 else if (vdm.sign == 0) {
 282  sqrt_copy:
 283                         vdp = &vdm;
 284                         ret = 0;
 285                 } else {
 286  sqrt_invalid:
 287                         vdp = &vfp_double_default_qnan;
 288                         ret = FPSCR_IOC;
 289                 }
 290                 vfp_put_double(dd, vfp_double_pack(vdp));
 291                 return ret;
 292         }
 293
 294         /*
 295          * sqrt(+/- 0) == +/- 0
 296          */
 297         if (tm & VFP_ZERO)
 298                 goto sqrt_copy;
 299
 300         /*
 301          * Normalise a denormalised number
 302          */
 303         if (tm & VFP_DENORMAL)
 304                 vfp_double_normalise_denormal(&vdm);
 305
 306         /*
 307          * sqrt(<0) = invalid
 308          */
 309         if (vdm.sign)
 310                 goto sqrt_invalid;
 311
 312         vfp_double_dump("sqrt", &vdm);
 313
 314         /*
 315          * Estimate the square root.
 316          */
 317         vdd.sign = 0;
 318         vdd.exponent = ((vdm.exponent - 1023) >> 1) + 1023;
 319         vdd.significand = (u64)vfp_estimate_sqrt_significand(vdm.exponent, vdm.significand >> 32) << 31;
 320
 321         vfp_double_dump("sqrt estimate1", &vdd);
 322
 323         vdm.significand >>= 1 + (vdm.exponent & 1);
 324         vdd.significand += 2 + vfp_estimate_div128to64(vdm.significand, 0, vdd.significand);
 325
 326         vfp_double_dump("sqrt estimate2", &vdd);
 327
 328         /*
 329          * And now adjust.
 330          */
 331         if ((vdd.significand & VFP_DOUBLE_LOW_BITS_MASK) <= 5) {
 332                 if (vdd.significand < 2) {
 333                         vdd.significand = ~0ULL;
 334                 } else {
 335                         u64 termh, terml, remh, reml;
 336                         vdm.significand <<= 2;
 337                         mul64to128(&termh, &terml, vdd.significand, vdd.significand);
 338                         sub128(&remh, &reml, vdm.significand, 0, termh, terml);
 339                         while ((s64)remh < 0) {
 340                                 vdd.significand -= 1;
 341                                 shift64left(&termh, &terml, vdd.significand);
 342                                 terml |= 1;
 343                                 add128(&remh, &reml, remh, reml, termh, terml);
 344                         }
 345                         vdd.significand |= (remh | reml) != 0;
 346                 }
 347         }
 348         vdd.significand = vfp_shiftright64jamming(vdd.significand, 1);
 349
 350         return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fsqrt");
 351 }
 352
 353 /*
 354  * Equal        := ZC
 355  * Less than    := N
 356  * Greater than := C
 357  * Unordered    := CV
 358  */
 359 static u32 vfp_compare(int dd, int signal_on_qnan, int dm, u32 fpscr)
 360 {
 361         s64 d, m;
 362         u32 ret = 0;
 363
 364         m = vfp_get_double(dm);
 365         if (vfp_double_packed_exponent(m) == 2047 && vfp_double_packed_mantissa(m)) {
 366                 ret |= FPSCR_C | FPSCR_V;
 367                 if (signal_on_qnan || !(vfp_double_packed_mantissa(m) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
 368                         /*
 369                          * Signalling NaN, or signalling on quiet NaN
 370                          */
 371                         ret |= FPSCR_IOC;
 372         }
 373
 374         d = vfp_get_double(dd);
 375         if (vfp_double_packed_exponent(d) == 2047 && vfp_double_packed_mantissa(d)) {
 376                 ret |= FPSCR_C | FPSCR_V;
 377                 if (signal_on_qnan || !(vfp_double_packed_mantissa(d) & (1ULL << (VFP_DOUBLE_MANTISSA_BITS - 1))))
 378                         /*
 379                          * Signalling NaN, or signalling on quiet NaN
 380                          */
 381                         ret |= FPSCR_IOC;
 382         }
 383
 384         if (ret == 0) {
 385                 if (d == m || vfp_double_packed_abs(d | m) == 0) {
 386                         /*
 387                          * equal
 388                          */
 389                         ret |= FPSCR_Z | FPSCR_C;
 390                 } else if (vfp_double_packed_sign(d ^ m)) {
 391                         /*
 392                          * different signs
 393                          */
 394                         if (vfp_double_packed_sign(d))
 395                                 /*
 396                                  * d is negative, so d < m
 397                                  */
 398                                 ret |= FPSCR_N;
 399                         else
 400                                 /*
 401                                  * d is positive, so d > m
 402                                  */
 403                                 ret |= FPSCR_C;
 404                 } else if ((vfp_double_packed_sign(d) != 0) ^ (d < m)) {
 405                         /*
 406                          * d < m
 407                          */
 408                         ret |= FPSCR_N;
 409                 } else if ((vfp_double_packed_sign(d) != 0) ^ (d > m)) {
 410                         /*
 411                          * d > m
 412                          */
 413                         ret |= FPSCR_C;
 414                 }
 415         }
 416
 417         return ret;
 418 }
 419
 420 static u32 vfp_double_fcmp(int dd, int unused, int dm, u32 fpscr)
 421 {
 422         return vfp_compare(dd, 0, dm, fpscr);
 423 }
 424
 425 static u32 vfp_double_fcmpe(int dd, int unused, int dm, u32 fpscr)
 426 {
 427         return vfp_compare(dd, 1, dm, fpscr);
 428 }
 429
 430 static u32 vfp_double_fcmpz(int dd, int unused, int dm, u32 fpscr)
 431 {
 432         return vfp_compare(dd, 0, VFP_REG_ZERO, fpscr);
 433 }
 434
 435 static u32 vfp_double_fcmpez(int dd, int unused, int dm, u32 fpscr)
 436 {
 437         return vfp_compare(dd, 1, VFP_REG_ZERO, fpscr);
 438 }
 439
 440 static u32 vfp_double_fcvts(int sd, int unused, int dm, u32 fpscr)
 441 {
 442         struct vfp_double vdm;
 443         struct vfp_single vsd;
 444         int tm;
 445         u32 exceptions = 0;
 446
 447         vfp_double_unpack(&vdm, vfp_get_double(dm));
 448
 449         tm = vfp_double_type(&vdm);
 450
 451         /*
 452          * If we have a signalling NaN, signal invalid operation.
 453          */
 454         if (tm == VFP_SNAN)
 455                 exceptions = FPSCR_IOC;
 456
 457         if (tm & VFP_DENORMAL)
 458                 vfp_double_normalise_denormal(&vdm);
 459
 460         vsd.sign = vdm.sign;
 461         vsd.significand = vfp_hi64to32jamming(vdm.significand);
 462
 463         /*
 464          * If we have an infinity or a NaN, the exponent must be 255
 465          */
 466         if (tm & (VFP_INFINITY|VFP_NAN)) {
 467                 vsd.exponent = 255;
 468                 if (tm & VFP_NAN)
 469                         vsd.significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
 470                 goto pack_nan;
 471         } else if (tm & VFP_ZERO)
 472                 vsd.exponent = 0;
 473         else
 474                 vsd.exponent = vdm.exponent - (1023 - 127);
 475
 476         return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fcvts");
 477
 478  pack_nan:
 479         vfp_put_float(sd, vfp_single_pack(&vsd));
 480         return exceptions;
 481 }
 482
 483 static u32 vfp_double_fuito(int dd, int unused, int dm, u32 fpscr)
 484 {
 485         struct vfp_double vdm;
 486         u32 m = vfp_get_float(dm);
 487
 488         vdm.sign = 0;
 489         vdm.exponent = 1023 + 63 - 1;
 490         vdm.significand = (u64)m;
 491
 492         return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fuito");
 493 }
 494
 495 static u32 vfp_double_fsito(int dd, int unused, int dm, u32 fpscr)
 496 {
 497         struct vfp_double vdm;
 498         u32 m = vfp_get_float(dm);
 499
 500         vdm.sign = (m & 0x80000000) >> 16;
 501         vdm.exponent = 1023 + 63 - 1;
 502         vdm.significand = vdm.sign ? -m : m;
 503
 504         return vfp_double_normaliseround(dd, &vdm, fpscr, 0, "fsito");
 505 }
 506
 507 static u32 vfp_double_ftoui(int sd, int unused, int dm, u32 fpscr)
 508 {
 509         struct vfp_double vdm;
 510         u32 d, exceptions = 0;
 511         int rmode = fpscr & FPSCR_RMODE_MASK;
 512         int tm;
 513
 514         vfp_double_unpack(&vdm, vfp_get_double(dm));
 515
 516         /*
 517          * Do we have a denormalised number?
 518          */
 519         tm = vfp_double_type(&vdm);
 520         if (tm & VFP_DENORMAL)
 521                 exceptions |= FPSCR_IDC;
 522
 523         if (tm & VFP_NAN)
 524                 vdm.sign = 0;
 525
 526         if (vdm.exponent >= 1023 + 32) {
 527                 d = vdm.sign ? 0 : 0xffffffff;
 528                 exceptions = FPSCR_IOC;
 529         } else if (vdm.exponent >= 1023 - 1) {
 530                 int shift = 1023 + 63 - vdm.exponent;
 531                 u64 rem, incr = 0;
 532
 533                 /*
 534                  * 2^0 <= m < 2^32-2^8
 535                  */
 536                 d = (vdm.significand << 1) >> shift;
 537                 rem = vdm.significand << (65 - shift);
 538
 539                 if (rmode == FPSCR_ROUND_NEAREST) {
 540                         incr = 0x8000000000000000ULL;
 541                         if ((d & 1) == 0)
 542                                 incr -= 1;
 543                 } else if (rmode == FPSCR_ROUND_TOZERO) {
 544                         incr = 0;
 545                 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vdm.sign != 0)) {
 546                         incr = ~0ULL;
 547                 }
 548
 549                 if ((rem + incr) < rem) {
 550                         if (d < 0xffffffff)
 551                                 d += 1;
 552                         else
 553                                 exceptions |= FPSCR_IOC;
 554                 }
 555
 556                 if (d && vdm.sign) {
 557                         d = 0;
 558                         exceptions |= FPSCR_IOC;
 559                 } else if (rem)
 560                         exceptions |= FPSCR_IXC;
 561         } else {
 562                 d = 0;
 563                 if (vdm.exponent | vdm.significand) {
 564                         exceptions |= FPSCR_IXC;
 565                         if (rmode == FPSCR_ROUND_PLUSINF && vdm.sign == 0)
 566                                 d = 1;
 567                         else if (rmode == FPSCR_ROUND_MINUSINF && vdm.sign) {
 568                                 d = 0;
 569                                 exceptions |= FPSCR_IOC;
 570                         }
 571                 }
 572         }
 573
 574         pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
 575
 576         vfp_put_float(sd, d);
 577
 578         return exceptions;
 579 }
 580
 581 static u32 vfp_double_ftouiz(int sd, int unused, int dm, u32 fpscr)
 582 {
 583         return vfp_double_ftoui(sd, unused, dm, FPSCR_ROUND_TOZERO);
 584 }
 585
 586 static u32 vfp_double_ftosi(int sd, int unused, int dm, u32 fpscr)
 587 {
 588         struct vfp_double vdm;
 589         u32 d, exceptions = 0;
 590         int rmode = fpscr & FPSCR_RMODE_MASK;
 591
 592         vfp_double_unpack(&vdm, vfp_get_double(dm));
 593         vfp_double_dump("VDM", &vdm);
 594
 595         /*
 596          * Do we have denormalised number?
 597          */
 598         if (vfp_double_type(&vdm) & VFP_DENORMAL)
 599                 exceptions |= FPSCR_IDC;
 600
 601         if (vdm.exponent >= 1023 + 32) {
 602                 d = 0x7fffffff;
 603                 if (vdm.sign)
 604                         d = ~d;
 605                 exceptions |= FPSCR_IOC;
 606         } else if (vdm.exponent >= 1023 - 1) {
 607                 int shift = 1023 + 63 - vdm.exponent;   /* 58 */
 608                 u64 rem, incr = 0;
 609
 610                 d = (vdm.significand << 1) >> shift;
 611                 rem = vdm.significand << (65 - shift);
 612
 613                 if (rmode == FPSCR_ROUND_NEAREST) {
 614                         incr = 0x8000000000000000ULL;
 615                         if ((d & 1) == 0)
 616                                 incr -= 1;
 617                 } else if (rmode == FPSCR_ROUND_TOZERO) {
 618                         incr = 0;
 619                 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vdm.sign != 0)) {
 620                         incr = ~0ULL;
 621                 }
 622
 623                 if ((rem + incr) < rem && d < 0xffffffff)
 624                         d += 1;
 625                 if (d > 0x7fffffff + (vdm.sign != 0)) {
 626                         d = 0x7fffffff + (vdm.sign != 0);
 627                         exceptions |= FPSCR_IOC;
 628                 } else if (rem)
 629                         exceptions |= FPSCR_IXC;
 630
 631                 if (vdm.sign)
 632                         d = -d;
 633         } else {
 634                 d = 0;
 635                 if (vdm.exponent | vdm.significand) {
 636                         exceptions |= FPSCR_IXC;
 637                         if (rmode == FPSCR_ROUND_PLUSINF && vdm.sign == 0)
 638                                 d = 1;
 639                         else if (rmode == FPSCR_ROUND_MINUSINF && vdm.sign)
 640                                 d = -1;
 641                 }
 642         }
 643
 644         pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
 645
 646         vfp_put_float(sd, (s32)d);
 647
 648         return exceptions;
 649 }
 650
 651 static u32 vfp_double_ftosiz(int dd, int unused, int dm, u32 fpscr)
 652 {
 653         return vfp_double_ftosi(dd, unused, dm, FPSCR_ROUND_TOZERO);
 654 }
 655
 656
 657 static u32 (* const fop_extfns[32])(int dd, int unused, int dm, u32 fpscr) = {
 658         [FEXT_TO_IDX(FEXT_FCPY)]        = vfp_double_fcpy,
 659         [FEXT_TO_IDX(FEXT_FABS)]        = vfp_double_fabs,
 660         [FEXT_TO_IDX(FEXT_FNEG)]        = vfp_double_fneg,
 661         [FEXT_TO_IDX(FEXT_FSQRT)]       = vfp_double_fsqrt,
 662         [FEXT_TO_IDX(FEXT_FCMP)]        = vfp_double_fcmp,
 663         [FEXT_TO_IDX(FEXT_FCMPE)]       = vfp_double_fcmpe,
 664         [FEXT_TO_IDX(FEXT_FCMPZ)]       = vfp_double_fcmpz,
 665         [FEXT_TO_IDX(FEXT_FCMPEZ)]      = vfp_double_fcmpez,
 666         [FEXT_TO_IDX(FEXT_FCVT)]        = vfp_double_fcvts,
 667         [FEXT_TO_IDX(FEXT_FUITO)]       = vfp_double_fuito,
 668         [FEXT_TO_IDX(FEXT_FSITO)]       = vfp_double_fsito,
 669         [FEXT_TO_IDX(FEXT_FTOUI)]       = vfp_double_ftoui,
 670         [FEXT_TO_IDX(FEXT_FTOUIZ)]      = vfp_double_ftouiz,
 671         [FEXT_TO_IDX(FEXT_FTOSI)]       = vfp_double_ftosi,
 672         [FEXT_TO_IDX(FEXT_FTOSIZ)]      = vfp_double_ftosiz,
 673 };
 674
 675
 676
 677
 678 static u32
 679 vfp_double_fadd_nonnumber(struct vfp_double *vdd, struct vfp_double *vdn,
 680                           struct vfp_double *vdm, u32 fpscr)
 681 {
 682         struct vfp_double *vdp;
 683         u32 exceptions = 0;
 684         int tn, tm;
 685
 686         tn = vfp_double_type(vdn);
 687         tm = vfp_double_type(vdm);
 688
 689         if (tn & tm & VFP_INFINITY) {
 690                 /*
 691                  * Two infinities.  Are they different signs?
 692                  */
 693                 if (vdn->sign ^ vdm->sign) {
 694                         /*
 695                          * different signs -> invalid
 696                          */
 697                         exceptions = FPSCR_IOC;
 698                         vdp = &vfp_double_default_qnan;
 699                 } else {
 700                         /*
 701                          * same signs -> valid
 702                          */
 703                         vdp = vdn;
 704                 }
 705         } else if (tn & VFP_INFINITY && tm & VFP_NUMBER) {
 706                 /*
 707                  * One infinity and one number -> infinity
 708                  */
 709                 vdp = vdn;
 710         } else {
 711                 /*
 712                  * 'n' is a NaN of some type
 713                  */
 714                 return vfp_propagate_nan(vdd, vdn, vdm, fpscr);
 715         }
 716         *vdd = *vdp;
 717         return exceptions;
 718 }
 719
 720 static u32
 721 vfp_double_add(struct vfp_double *vdd, struct vfp_double *vdn,
 722                struct vfp_double *vdm, u32 fpscr)
 723 {
 724         u32 exp_diff;
 725         u64 m_sig;
 726
 727         if (vdn->significand & (1ULL << 63) ||
 728             vdm->significand & (1ULL << 63)) {
 729                 pr_info("VFP: bad FP values in %s\n", __func__);
 730                 vfp_double_dump("VDN", vdn);
 731                 vfp_double_dump("VDM", vdm);
 732         }
 733
 734         /*
 735          * Ensure that 'n' is the largest magnitude number.  Note that
 736          * if 'n' and 'm' have equal exponents, we do not swap them.
 737          * This ensures that NaN propagation works correctly.
 738          */
 739         if (vdn->exponent < vdm->exponent) {
 740                 struct vfp_double *t = vdn;
 741                 vdn = vdm;
 742                 vdm = t;
 743         }
 744
 745         /*
 746          * Is 'n' an infinity or a NaN?  Note that 'm' may be a number,
 747          * infinity or a NaN here.
 748          */
 749         if (vdn->exponent == 2047)
 750                 return vfp_double_fadd_nonnumber(vdd, vdn, vdm, fpscr);
 751
 752         /*
 753          * We have two proper numbers, where 'vdn' is the larger magnitude.
 754          *
 755          * Copy 'n' to 'd' before doing the arithmetic.
 756          */
 757         *vdd = *vdn;
 758
 759         /*
 760          * Align 'm' with the result.
 761          */
 762         exp_diff = vdn->exponent - vdm->exponent;
 763         m_sig = vfp_shiftright64jamming(vdm->significand, exp_diff);
 764
 765         /*
 766          * If the signs are different, we are really subtracting.
 767          */
 768         if (vdn->sign ^ vdm->sign) {
 769                 m_sig = vdn->significand - m_sig;
 770                 if ((s64)m_sig < 0) {
 771                         vdd->sign = vfp_sign_negate(vdd->sign);
 772                         m_sig = -m_sig;
 773                 }
 774         } else {
 775                 m_sig += vdn->significand;
 776         }
 777         vdd->significand = m_sig;
 778
 779         return 0;
 780 }
 781
 782 static u32
 783 vfp_double_multiply(struct vfp_double *vdd, struct vfp_double *vdn,
 784                     struct vfp_double *vdm, u32 fpscr)
 785 {
 786         vfp_double_dump("VDN", vdn);
 787         vfp_double_dump("VDM", vdm);
 788
 789         /*
 790          * Ensure that 'n' is the largest magnitude number.  Note that
 791          * if 'n' and 'm' have equal exponents, we do not swap them.
 792          * This ensures that NaN propagation works correctly.
 793          */
 794         if (vdn->exponent < vdm->exponent) {
 795                 struct vfp_double *t = vdn;
 796                 vdn = vdm;
 797                 vdm = t;
 798                 pr_debug("VFP: swapping M <-> N\n");
 799         }
 800
 801         vdd->sign = vdn->sign ^ vdm->sign;
 802
 803         /*
 804          * If 'n' is an infinity or NaN, handle it.  'm' may be anything.
 805          */
 806         if (vdn->exponent == 2047) {
 807                 if (vdn->significand || (vdm->exponent == 2047 && vdm->significand))
 808                         return vfp_propagate_nan(vdd, vdn, vdm, fpscr);
 809                 if ((vdm->exponent | vdm->significand) == 0) {
 810                         *vdd = vfp_double_default_qnan;
 811                         return FPSCR_IOC;
 812                 }
 813                 vdd->exponent = vdn->exponent;
 814                 vdd->significand = 0;
 815                 return 0;
 816         }
 817
 818         /*
 819          * If 'm' is zero, the result is always zero.  In this case,
 820          * 'n' may be zero or a number, but it doesn't matter which.
 821          */
 822         if ((vdm->exponent | vdm->significand) == 0) {
 823                 vdd->exponent = 0;
 824                 vdd->significand = 0;
 825                 return 0;
 826         }
 827
 828         /*
 829          * We add 2 to the destination exponent for the same reason
 830          * as the addition case - though this time we have +1 from
 831          * each input operand.
 832          */
 833         vdd->exponent = vdn->exponent + vdm->exponent - 1023 + 2;
 834         vdd->significand = vfp_hi64multiply64(vdn->significand, vdm->significand);
 835
 836         vfp_double_dump("VDD", vdd);
 837         return 0;
 838 }
 839
 840 #define NEG_MULTIPLY    (1 << 0)
 841 #define NEG_SUBTRACT    (1 << 1)
 842
 843 static u32
 844 vfp_double_multiply_accumulate(int dd, int dn, int dm, u32 fpscr, u32 negate, char *func)
 845 {
 846         struct vfp_double vdd, vdp, vdn, vdm;
 847         u32 exceptions;
 848
 849         vfp_double_unpack(&vdn, vfp_get_double(dn));
 850         if (vdn.exponent == 0 && vdn.significand)
 851                 vfp_double_normalise_denormal(&vdn);
 852
 853         vfp_double_unpack(&vdm, vfp_get_double(dm));
 854         if (vdm.exponent == 0 && vdm.significand)
 855                 vfp_double_normalise_denormal(&vdm);
 856
 857         exceptions = vfp_double_multiply(&vdp, &vdn, &vdm, fpscr);
 858         if (negate & NEG_MULTIPLY)
 859                 vdp.sign = vfp_sign_negate(vdp.sign);
 860
 861         vfp_double_unpack(&vdn, vfp_get_double(dd));
 862         if (negate & NEG_SUBTRACT)
 863                 vdn.sign = vfp_sign_negate(vdn.sign);
 864
 865         exceptions |= vfp_double_add(&vdd, &vdn, &vdp, fpscr);
 866
 867         return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, func);
 868 }
 869
 870 /*
 871  * Standard operations
 872  */
 873
 874 /*
 875  * sd = sd + (sn * sm)
 876  */
 877 static u32 vfp_double_fmac(int dd, int dn, int dm, u32 fpscr)
 878 {
 879         return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, 0, "fmac");
 880 }
 881
 882 /*
 883  * sd = sd - (sn * sm)
 884  */
 885 static u32 vfp_double_fnmac(int dd, int dn, int dm, u32 fpscr)
 886 {
 887         return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_MULTIPLY, "fnmac");
 888 }
 889
 890 /*
 891  * sd = -sd + (sn * sm)
 892  */
 893 static u32 vfp_double_fmsc(int dd, int dn, int dm, u32 fpscr)
 894 {
 895         return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT, "fmsc");
 896 }
 897
 898 /*
 899  * sd = -sd - (sn * sm)
 900  */
 901 static u32 vfp_double_fnmsc(int dd, int dn, int dm, u32 fpscr)
 902 {
 903         return vfp_double_multiply_accumulate(dd, dn, dm, fpscr, NEG_SUBTRACT | NEG_MULTIPLY, "fnmsc");
 904 }
 905
 906 /*
 907  * sd = sn * sm
 908  */
 909 static u32 vfp_double_fmul(int dd, int dn, int dm, u32 fpscr)
 910 {
 911         struct vfp_double vdd, vdn, vdm;
 912         u32 exceptions;
 913
 914         vfp_double_unpack(&vdn, vfp_get_double(dn));
 915         if (vdn.exponent == 0 && vdn.significand)
 916                 vfp_double_normalise_denormal(&vdn);
 917
 918         vfp_double_unpack(&vdm, vfp_get_double(dm));
 919         if (vdm.exponent == 0 && vdm.significand)
 920                 vfp_double_normalise_denormal(&vdm);
 921
 922         exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
 923         return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fmul");
 924 }
 925
 926 /*
 927  * sd = -(sn * sm)
 928  */
 929 static u32 vfp_double_fnmul(int dd, int dn, int dm, u32 fpscr)
 930 {
 931         struct vfp_double vdd, vdn, vdm;
 932         u32 exceptions;
 933
 934         vfp_double_unpack(&vdn, vfp_get_double(dn));
 935         if (vdn.exponent == 0 && vdn.significand)
 936                 vfp_double_normalise_denormal(&vdn);
 937
 938         vfp_double_unpack(&vdm, vfp_get_double(dm));
 939         if (vdm.exponent == 0 && vdm.significand)
 940                 vfp_double_normalise_denormal(&vdm);
 941
 942         exceptions = vfp_double_multiply(&vdd, &vdn, &vdm, fpscr);
 943         vdd.sign = vfp_sign_negate(vdd.sign);
 944
 945         return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fnmul");
 946 }
 947
 948 /*
 949  * sd = sn + sm
 950  */
 951 static u32 vfp_double_fadd(int dd, int dn, int dm, u32 fpscr)
 952 {
 953         struct vfp_double vdd, vdn, vdm;
 954         u32 exceptions;
 955
 956         vfp_double_unpack(&vdn, vfp_get_double(dn));
 957         if (vdn.exponent == 0 && vdn.significand)
 958                 vfp_double_normalise_denormal(&vdn);
 959
 960         vfp_double_unpack(&vdm, vfp_get_double(dm));
 961         if (vdm.exponent == 0 && vdm.significand)
 962                 vfp_double_normalise_denormal(&vdm);
 963
 964         exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
 965
 966         return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fadd");
 967 }
 968
 969 /*
 970  * sd = sn - sm
 971  */
 972 static u32 vfp_double_fsub(int dd, int dn, int dm, u32 fpscr)
 973 {
 974         struct vfp_double vdd, vdn, vdm;
 975         u32 exceptions;
 976
 977         vfp_double_unpack(&vdn, vfp_get_double(dn));
 978         if (vdn.exponent == 0 && vdn.significand)
 979                 vfp_double_normalise_denormal(&vdn);
 980
 981         vfp_double_unpack(&vdm, vfp_get_double(dm));
 982         if (vdm.exponent == 0 && vdm.significand)
 983                 vfp_double_normalise_denormal(&vdm);
 984
 985         /*
 986          * Subtraction is like addition, but with a negated operand.
 987          */
 988         vdm.sign = vfp_sign_negate(vdm.sign);
 989
 990         exceptions = vfp_double_add(&vdd, &vdn, &vdm, fpscr);
 991
 992         return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fsub");
 993 }
 994
 995 /*
 996  * sd = sn / sm
 997  */
 998 static u32 vfp_double_fdiv(int dd, int dn, int dm, u32 fpscr)
 999 {
1000         struct vfp_double vdd, vdn, vdm;
1001         u32 exceptions = 0;
1002         int tm, tn;
1003
1004         vfp_double_unpack(&vdn, vfp_get_double(dn));
1005         vfp_double_unpack(&vdm, vfp_get_double(dm));
1006
1007         vdd.sign = vdn.sign ^ vdm.sign;
1008
1009         tn = vfp_double_type(&vdn);
1010         tm = vfp_double_type(&vdm);
1011
1012         /*
1013          * Is n a NAN?
1014          */
1015         if (tn & VFP_NAN)
1016                 goto vdn_nan;
1017
1018         /*
1019          * Is m a NAN?
1020          */
1021         if (tm & VFP_NAN)
1022                 goto vdm_nan;
1023
1024         /*
1025          * If n and m are infinity, the result is invalid
1026          * If n and m are zero, the result is invalid
1027          */
1028         if (tm & tn & (VFP_INFINITY|VFP_ZERO))
1029                 goto invalid;
1030
1031         /*
1032          * If n is infinity, the result is infinity
1033          */
1034         if (tn & VFP_INFINITY)
1035                 goto infinity;
1036
1037         /*
1038          * If m is zero, raise div0 exceptions
1039          */
1040         if (tm & VFP_ZERO)
1041                 goto divzero;
1042
1043         /*
1044          * If m is infinity, or n is zero, the result is zero
1045          */
1046         if (tm & VFP_INFINITY || tn & VFP_ZERO)
1047                 goto zero;
1048
1049         if (tn & VFP_DENORMAL)
1050                 vfp_double_normalise_denormal(&vdn);
1051         if (tm & VFP_DENORMAL)
1052                 vfp_double_normalise_denormal(&vdm);
1053
1054         /*
1055          * Ok, we have two numbers, we can perform division.
1056          */
1057         vdd.exponent = vdn.exponent - vdm.exponent + 1023 - 1;
1058         vdm.significand <<= 1;
1059         if (vdm.significand <= (2 * vdn.significand)) {
1060                 vdn.significand >>= 1;
1061                 vdd.exponent++;
1062         }
1063         vdd.significand = vfp_estimate_div128to64(vdn.significand, 0, vdm.significand);
1064         if ((vdd.significand & 0x1ff) <= 2) {
1065                 u64 termh, terml, remh, reml;
1066                 mul64to128(&termh, &terml, vdm.significand, vdd.significand);
1067                 sub128(&remh, &reml, vdn.significand, 0, termh, terml);
1068                 while ((s64)remh < 0) {
1069                         vdd.significand -= 1;
1070                         add128(&remh, &reml, remh, reml, 0, vdm.significand);
1071                 }
1072                 vdd.significand |= (reml != 0);
1073         }
1074         return vfp_double_normaliseround(dd, &vdd, fpscr, 0, "fdiv");
1075
1076  vdn_nan:
1077         exceptions = vfp_propagate_nan(&vdd, &vdn, &vdm, fpscr);
1078  pack:
1079         vfp_put_double(dd, vfp_double_pack(&vdd));
1080         return exceptions;
1081
1082  vdm_nan:
1083         exceptions = vfp_propagate_nan(&vdd, &vdm, &vdn, fpscr);
1084         goto pack;
1085
1086  zero:
1087         vdd.exponent = 0;
1088         vdd.significand = 0;
1089         goto pack;
1090
1091  divzero:
1092         exceptions = FPSCR_DZC;
1093  infinity:
1094         vdd.exponent = 2047;
1095         vdd.significand = 0;
1096         goto pack;
1097
1098  invalid:
1099         vfp_put_double(dd, vfp_double_pack(&vfp_double_default_qnan));
1100         return FPSCR_IOC;
1101 }
1102
1103 static u32 (* const fop_fns[16])(int dd, int dn, int dm, u32 fpscr) = {
1104         [FOP_TO_IDX(FOP_FMAC)]  = vfp_double_fmac,
1105         [FOP_TO_IDX(FOP_FNMAC)] = vfp_double_fnmac,
1106         [FOP_TO_IDX(FOP_FMSC)]  = vfp_double_fmsc,
1107         [FOP_TO_IDX(FOP_FNMSC)] = vfp_double_fnmsc,
1108         [FOP_TO_IDX(FOP_FMUL)]  = vfp_double_fmul,
1109         [FOP_TO_IDX(FOP_FNMUL)] = vfp_double_fnmul,
1110         [FOP_TO_IDX(FOP_FADD)]  = vfp_double_fadd,
1111         [FOP_TO_IDX(FOP_FSUB)]  = vfp_double_fsub,
1112         [FOP_TO_IDX(FOP_FDIV)]  = vfp_double_fdiv,
1113 };
1114
1115 #define FREG_BANK(x)    ((x) & 0x0c)
1116 #define FREG_IDX(x)     ((x) & 3)
1117
1118 u32 vfp_double_cpdo(u32 inst, u32 fpscr)
1119 {
1120         u32 op = inst & FOP_MASK;
1121         u32 exceptions = 0;
1122         unsigned int dd = vfp_get_sd(inst);
1123         unsigned int dn = vfp_get_sn(inst);
1124         unsigned int dm = vfp_get_sm(inst);
1125         unsigned int vecitr, veclen, vecstride;
1126         u32 (*fop)(int, int, s32, u32);
1127
1128         veclen = fpscr & FPSCR_LENGTH_MASK;
1129         vecstride = (1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK)) * 2;
1130
1131         /*
1132          * If destination bank is zero, vector length is always '1'.
1133          * ARM DDI0100F C5.1.3, C5.3.2.
1134          */
1135         if (FREG_BANK(dd) == 0)
1136                 veclen = 0;
1137
1138         pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride,
1139                  (veclen >> FPSCR_LENGTH_BIT) + 1);
1140
1141         fop = (op == FOP_EXT) ? fop_extfns[dn] : fop_fns[FOP_TO_IDX(op)];
1142         if (!fop)
1143                 goto invalid;
1144
1145         for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) {
1146                 u32 except;
1147
1148                 if (op == FOP_EXT)
1149                         pr_debug("VFP: itr%d (d%u.%u) = op[%u] (d%u.%u)\n",
1150                                  vecitr >> FPSCR_LENGTH_BIT,
1151                                  dd >> 1, dd & 1, dn,
1152                                  dm >> 1, dm & 1);
1153                 else
1154                         pr_debug("VFP: itr%d (d%u.%u) = (d%u.%u) op[%u] (d%u.%u)\n",
1155                                  vecitr >> FPSCR_LENGTH_BIT,
1156                                  dd >> 1, dd & 1,
1157                                  dn >> 1, dn & 1,
1158                                  FOP_TO_IDX(op),
1159                                  dm >> 1, dm & 1);
1160
1161                 except = fop(dd, dn, dm, fpscr);
1162                 pr_debug("VFP: itr%d: exceptions=%08x\n",
1163                          vecitr >> FPSCR_LENGTH_BIT, except);
1164
1165                 exceptions |= except;
1166
1167                 /*
1168                  * This ensures that comparisons only operate on scalars;
1169                  * comparisons always return with one FPSCR status bit set.
1170                  */
1171                 if (except & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
1172                         break;
1173
1174                 /*
1175                  * CHECK: It appears to be undefined whether we stop when
1176                  * we encounter an exception.  We continue.
1177                  */
1178
1179                 dd = FREG_BANK(dd) + ((FREG_IDX(dd) + vecstride) & 6);
1180                 dn = FREG_BANK(dn) + ((FREG_IDX(dn) + vecstride) & 6);
1181                 if (FREG_BANK(dm) != 0)
1182                         dm = FREG_BANK(dm) + ((FREG_IDX(dm) + vecstride) & 6);
1183         }
1184         return exceptions;
1185
1186  invalid:
1187         return ~0;
1188 }