fltcalc.c 42 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
/*
 * Project:     libFIRM
 * File name:   ir/tv/fltcalc.c
 * Purpose:
 * Author:
 * Modified by:
 * Created:     2003
 * CVS-ID:      $Id$
 * Copyright:   (c) 2003 Universitt Karlsruhe
 * Licence:     This file protected by GPL -  GNU GENERAL PUBLIC LICENSE.
11
12
 */

13

Boris Boesler's avatar
Boris Boesler committed
14
15
16
17
18
#ifdef HAVE_CONFIG_H
# include "config.h"
#endif


19
#include "fltcalc.h"
20
21
22
23
24
25
26
27
28
#include "strcalc.h"

#include <math.h>    /* need isnan() and isinf() (will be changed)*/
/* undef some reused constants defined by math.h */
#ifdef NAN
#  undef NAN
#endif

#include <inttypes.h>
29
30
31
#include <string.h>
#include <stdlib.h>
#include <stdio.h>
32
#include <assert.h>
Boris Boesler's avatar
Boris Boesler committed
33
#ifdef HAVE_ALLOCA_H
Boris Boesler's avatar
Boris Boesler committed
34
35
# include <alloca.h>
#endif
36

37
typedef uint32_t UINT32;
Michael Beck's avatar
Tarval:    
Michael Beck committed
38

39
#ifdef HAVE_LONG_DOUBLE
40
#ifdef WORDS_BIGENDIAN
41
42
43
typedef union {
  struct {
    UINT32 high;
44
45
    UINT32 mid;
    UINT32 low;
46
47
48
49
50
51
52
  } val;
  volatile long double d;
} value_t;
#else
typedef union {
  struct {
    UINT32 low;
53
54
    UINT32 mid;
    UINT32 high;
55
56
57
58
59
  } val;
  volatile long double d;
} value_t;
#endif
#else
60
#ifdef WORDS_BIGENDIAN
61
62
63
typedef union {
  struct {
    UINT32 high;
64
    UINT32 low;
65
66
67
68
69
70
71
  } val;
  volatile double d;
} value_t;
#else
typedef union {
  struct {
    UINT32 low;
72
    UINT32 high;
73
74
75
76
77
  } val;
  volatile double d;
} value_t;
#endif
#endif
78

Michael Beck's avatar
Tarval:    
Michael Beck committed
79

80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
/********
 * globals
 ********/
typedef enum {
  NORMAL,
  ZERO,
  SUBNORMAL,
  INF,
  NAN,
} value_class_t;

typedef struct {
  char  exponent_size;
  char  mantissa_size;
  value_class_t  class;
} descriptor_t;

#define CLEAR_BUFFER(buffer) memset(buffer, 0, CALC_BUFFER_SIZE)

/* because variable sized structs are impossible, the internal
 * value is represented as a pseudo-struct char array, addressed
 * by macros
 * struct {
 *   char sign;             // 0 for positive, 1 for negative
 *   char exp[VALUE_SIZE];
 *   char mant[VALUE_SIZE];
 *   descriptor_t desc;
 * };
 */
#define _sign(a) (((char*)a)[SIGN_POS])
#define _exp(a) (&((char*)a)[EXPONENT_POS])
#define _mant(a) (&((char*)a)[MANTISSA_POS])
#define _desc(a) (*(descriptor_t *)&((char*)a)[DESCRIPTOR_POS])
Michael Beck's avatar
Tarval:    
Michael Beck committed
113

114
115
116
#define _save_result(x) memcpy((x), sc_get_buffer(), VALUE_SIZE)
#define _shift_right(x, y, b) sc_shr((x), (y), VALUE_SIZE*4, 0, (b))
#define _shift_left(x, y, b) sc_shl((x), (y), VALUE_SIZE*4, 0, (b))
Michael Beck's avatar
Tarval:    
Michael Beck committed
117

118
119
120
121
#define FC_DEFINE1(code) char* fc_##code(const void *a, void *result)                          \
                   {                                                                           \
                     return _calc((const char*)a, NULL, FC_##code, (char*)result);             \
                   }
Michael Beck's avatar
Tarval:    
Michael Beck committed
122

123
124
125
126
127
128
129
130
131
132
133
#define FC_DEFINE2(code) char* fc_##code(const void *a, const void *b, void *result)           \
                   {                                                                           \
                     return _calc((const char*)a, (const char*)b, FC_##code, (char*)result);   \
                   }

#define FUNC_PTR(code) fc_##code

#if FLTCALC_DEBUG
#  define DEBUGPRINTF(x) printf x
#else
#  define DEBUGPRINTF(x) ((void)0)
Michael Beck's avatar
Tarval:    
Michael Beck committed
134
135
#endif

136
137
138
139
140
#if FLTCALC_TRACE_CALC
#  define TRACEPRINTF(x) printf x
#else
#  define TRACEPRINTF(x) ((void)0)
#endif
141

142
static char *calc_buffer = NULL;
143

144
145
146
147
148
149
150
151
static fc_rounding_mode_t ROUNDING_MODE;

static int CALC_BUFFER_SIZE;
static int VALUE_SIZE;
static int SIGN_POS;
static int EXPONENT_POS;
static int MANTISSA_POS;
static int DESCRIPTOR_POS;
152

153
static int max_precision;
154
/********
155
 * private functions
156
 ********/
157
158
#if 0
static void _fail_char(const char *str, unsigned int len, int pos)
159
{
160
161
162
163
164
165
166
167
  if (*(str+pos))
    printf("ERROR: Unexpected character '%c'\n", *(str + pos));
  else
    printf("ERROR: Unexpected end of string\n");
  while (len-- && *str) printf("%c", *str++); printf("\n");
  while (pos--) printf(" "); printf("^\n");
  /* the front end has to to check constant strings */
  exit(-1);
168
}
169
#endif
170

171
/* pack machine-like */
172
static char* _pack(const char *int_float, char *packed)
173
{
174
175
176
  char *shift_val;
  char *temp;
  char *val_buffer;
177

178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
  temp = alloca(VALUE_SIZE);
  shift_val = alloca(VALUE_SIZE);

  switch (_desc(int_float).class) {
    case NAN:
      val_buffer = alloca(CALC_BUFFER_SIZE);
      fc_get_qnan(_desc(int_float).exponent_size, _desc(int_float).mantissa_size, val_buffer);
      int_float = val_buffer;
      break;

    case INF:
      val_buffer = alloca(CALC_BUFFER_SIZE);
      fc_get_plusinf(_desc(int_float).exponent_size, _desc(int_float).mantissa_size, val_buffer);
      _sign(val_buffer) = _sign(int_float);
      int_float = val_buffer;
      break;

    default:
      break;
  }
  /* pack sign */
  sc_val_from_ulong(_sign(int_float), temp);

  sc_val_from_ulong(_desc(int_float).exponent_size + _desc(int_float).mantissa_size, NULL);
  _shift_left(temp, sc_get_buffer(), packed);

  /* extract exponent */
  sc_val_from_ulong(_desc(int_float).mantissa_size, shift_val);

  _shift_left(_exp(int_float), shift_val, temp);

  sc_or(temp, packed, packed);

  /* extract mantissa */
  /* remove 2 rounding bits */
  sc_val_from_ulong(2, shift_val);
  _shift_right(_mant(int_float), shift_val, temp);

  /* remove leading 1 (or 0 if denormalized) */
  sc_max_from_bits(_desc(int_float).mantissa_size, 0, shift_val); /* all mantissa bits are 1's */
  sc_and(temp, shift_val, temp);

  /* save result */
  sc_or(temp, packed, packed);

  return packed;
224
225
}

226
char* _normalize(const char *in_val, char *out_val, int sticky)
227
{
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
  int hsb;
  char lsb, guard, round, round_dir = 0;
  char *temp;

  temp = alloca(VALUE_SIZE);

  /* +2: save two rounding bits at the end */
  hsb = 2 + _desc(in_val).mantissa_size - sc_get_highest_set_bit(_mant(in_val)) - 1;

  if (in_val != out_val)
  {
    _sign(out_val) = _sign(in_val);
    memcpy(&_desc(out_val), &_desc(in_val), sizeof(descriptor_t));
  }

  _desc(out_val).class = NORMAL;

  /* mantissa all zeroes, so zero exponent (because of explicit one)*/
  if (hsb == 2 + _desc(in_val).mantissa_size)
  {
    sc_val_from_ulong(0, _exp(out_val));
    hsb = -1;
  }

252
  /* shift the first 1 into the left of the radix point (i.e. hsb == -1) */
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
  if (hsb < -1)
  {
    /* shift right */
    sc_val_from_ulong(-hsb-1, temp);

    _shift_right(_mant(in_val), temp, _mant(out_val));

    /* remember if some bits were shifted away */
    if (!sticky) sticky = sc_had_carry();

    sc_add(_exp(in_val), temp, _exp(out_val));
  }
  else if (hsb > -1)
  {
    /* shift left */
    sc_val_from_ulong(hsb+1, temp);

    _shift_left(_mant(in_val), temp, _mant(out_val));

    sc_sub(_exp(in_val), temp, _exp(out_val));
  }

  /* check for exponent underflow */
  if (sc_is_negative(_exp(out_val)) || sc_is_zero(_exp(out_val))) {
    DEBUGPRINTF(("Exponent underflow!\n"));
    /* exponent underflow */
    /* shift the mantissa right to have a zero exponent */
    sc_val_from_ulong(1, temp);
    sc_sub(temp, _exp(out_val), NULL);

    _shift_right(_mant(out_val), sc_get_buffer(), _mant(out_val));
    if (!sticky) sticky = sc_had_carry();
    /* denormalized means exponent of zero */
    sc_val_from_ulong(0, _exp(out_val));

    _desc(out_val).class = SUBNORMAL;
  }

  /* perform rounding by adding a value that clears the guard bit and the round bit
   * and either causes a carry to round up or not */
  /* get the last 3 bits of the value */
  lsb = sc_sub_bits(_mant(out_val), _desc(out_val).mantissa_size + 2, 0) & 0x7;
  guard = (lsb&0x2)>>1;
  round = lsb&0x1;

  switch (ROUNDING_MODE)
  {
    case FC_TONEAREST:
      /* round to nearest representable value, if in doubt choose the version
       * with lsb == 0 */
      round_dir = guard && (sticky || round || lsb>>2);
      break;
    case FC_TOPOSITIVE:
      /* if positive: round to one if the exact value is bigger, else to zero */
      round_dir = (!_sign(out_val) && (guard || round || sticky));
      break;
    case FC_TONEGATIVE:
      /* if negative: round to one if the exact value is bigger, else to zero */
      round_dir = (_sign(out_val) && (guard || round || sticky));
      break;
    case FC_TOZERO:
      /* always round to 0 (chopping mode) */
      round_dir = 0;
      break;
  }
  DEBUGPRINTF(("Rounding (s%d, l%d, g%d, r%d, s%d) %s\n", _sign(out_val), lsb>>2, guard, round, sticky, (round_dir)?"up":"down"));

  if (round_dir == 1)
  {
    guard = (round^guard)<<1;
    lsb = !(round || guard)<<2 | guard | round;
  }
  else
  {
    lsb = -((guard<<1) | round);
  }

  /* add the rounded value */
  if (lsb != 0) {
    sc_val_from_long(lsb, temp);
    sc_add(_mant(out_val), temp, _mant(out_val));
  }

  /* could have rounded down to zero */
  if (sc_is_zero(_mant(out_val)) && (_desc(out_val).class == SUBNORMAL))
    _desc(out_val).class = ZERO;

  /* check for rounding overflow */
  hsb = 2 + _desc(out_val).mantissa_size - sc_get_highest_set_bit(_mant(out_val)) - 1;
  if ((_desc(out_val).class != SUBNORMAL) && (hsb < -1))
  {
    sc_val_from_ulong(1, temp);
    _shift_right(_mant(out_val), temp, _mant(out_val));

    sc_add(_exp(out_val), temp, _exp(out_val));
  }
  else if ((_desc(out_val).class == SUBNORMAL) && (hsb == -1))
  {
    /* overflow caused the matissa to be normal again,
     * so adapt the exponent accordingly */
    sc_val_from_ulong(1, temp);
    sc_add(_exp(out_val), temp, _exp(out_val));

    _desc(out_val).class = NORMAL;
  }
  /* no further rounding is needed, because rounding overflow means
   * the carry of the original rounding was propagated all the way
   * up to the bit left of the radix point. This implies the bits
   * to the right are all zeros (rounding is +1) */

  /* check for exponent overflow */
  sc_val_from_ulong((1 << _desc(out_val).exponent_size) - 1, temp);
  if (sc_comp(_exp(out_val), temp) != -1) {
    DEBUGPRINTF(("Exponent overflow!\n"));
    /* exponent overflow, reaction depends on rounding method:
     *
     * mode        | sign of value |  result
     *--------------------------------------------------------------
     * TO_NEAREST  |      +        |   +inf
     *             |      -        |   -inf
     *--------------------------------------------------------------
     * TO_POSITIVE |      +        |   +inf
     *             |      -        |   smallest representable value
     *--------------------------------------------------------------
     * TO_NEAGTIVE |      +        |   largest representable value
     *             |      -        |   -inf
     *--------------------------------------------------------------
     * TO_ZERO     |      +        |   largest representable value
     *             |      -        |   smallest representable value
     *--------------------------------------------------------------*/
    if (_sign(out_val) == 0)
    {
      /* value is positive */
      switch (ROUNDING_MODE) {
        case FC_TONEAREST:
        case FC_TOPOSITIVE:
          _desc(out_val).class = INF;
          break;

        case FC_TONEGATIVE:
        case FC_TOZERO:
          fc_get_max(_desc(out_val).exponent_size, _desc(out_val).mantissa_size, out_val);
      }
    } else {
      /* value is negative */
      switch (ROUNDING_MODE) {
        case FC_TONEAREST:
        case FC_TONEGATIVE:
          _desc(out_val).class = INF;
          break;

        case FC_TOPOSITIVE:
        case FC_TOZERO:
          fc_get_min(_desc(out_val).exponent_size, _desc(out_val).mantissa_size, out_val);
      }
    }
  }

  return out_val;
412
413
}

414
415
416
417
/*
 * calculate a + b, where a is the value with the bigger exponent
 */
static char* _add(const char* a, const char* b, char* result)
418
{
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
  char *temp;
  char *exp_diff;

  char sign;
  char sticky;

  if (_desc(a).class == NAN) {
    if (a != result) memcpy(result, a, CALC_BUFFER_SIZE);
    return result;
  }
  if (_desc(b).class == NAN) {
    if (b != result) memcpy(result, b, CALC_BUFFER_SIZE);
    return result;
  }

  /* make sure result has a descriptor */
  if (result != a && result != b)
    memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t));

  /* determine if this is an addition or subtraction */
  sign = _sign(a) ^ _sign(b);

  /* produce nan on inf - inf */
  if (sign && (_desc(a).class == INF) && (_desc(b).class == INF))
    return fc_get_qnan(_desc(a).exponent_size, _desc(b).mantissa_size, result);

  temp = alloca(VALUE_SIZE);
  exp_diff = alloca(VALUE_SIZE);

  /* get exponent difference */
  sc_sub(_exp(a), _exp(b), exp_diff);

  /* initially set sign to be the sign of a, special treatment of subtraction
   * when exponents are equal is required though.
   * Also special care about the sign is needed when the mantissas are equal
   * (+/- 0 ?) */
  if (sign && sc_val_to_long(exp_diff) == 0) {
    switch (sc_comp(_mant(a), _mant(b))) {
      case 1:  /* a > b */
        if (_sign(a)) _sign(result) = 1;  /* abs(a) is bigger and a is negative */
        else _sign(result) = 0;
        break;
      case 0:  /* a == b */
        if (ROUNDING_MODE == FC_TONEGATIVE)
          _sign(result) = 1;
        else
          _sign(result) = 0;
        break;
      case -1: /* a < b */
        if (_sign(b)) _sign(result) = 1; /* abs(b) is bigger and b is negative */
        else _sign(result) = 0;
        break;
      default:
        /* can't be reached */
        break;
    }
  } else {
    _sign(result) = _sign(a);
  }

  /* sign has been taken care of, check for special cases */
  if (_desc(a).class == ZERO) {
    if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, CALC_BUFFER_SIZE-SIGN_POS-1);
    return result;
  }
  if (_desc(b).class == ZERO) {
    if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-SIGN_POS-1);
    return result;
  }

  if (_desc(a).class == INF) {
    if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-SIGN_POS-1);
    return result;
  }
  if (_desc(b).class == INF) {
    if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, CALC_BUFFER_SIZE-SIGN_POS-1);
    return result;
  }

  /* shift the smaller value to the right to align the radix point */
  /* subnormals have their radix point shifted to the right,
   * take care of this first */
  if ((_desc(b).class == SUBNORMAL) && (_desc(a).class != SUBNORMAL))
  {
    sc_val_from_ulong(1, temp);
    sc_sub(exp_diff, temp, exp_diff);
  }

  _shift_right(_mant(b), exp_diff, temp);
  sticky = sc_had_carry();

  if (sticky && sign)
  {
    /* if subtracting a little more than the represented value or adding a little
     * more than the represented value to a negative value this, in addition to the
     * still set sticky bit, takes account of the 'little more' */
    char *temp1 = alloca(CALC_BUFFER_SIZE);
    sc_val_from_ulong(1, temp1);
    sc_add(temp, temp1, temp);
  }

  if (sign) {
    if (sc_comp(_mant(a), temp) == -1)
      sc_sub(temp, _mant(a), _mant(result));
    else
      sc_sub(_mant(a), temp, _mant(result));
  } else {
    sc_add(_mant(a), temp, _mant(result));
  }

  /* _normalize expects a 'normal' radix point, adding two subnormals
   * results in a subnormal radix point -> shifting before normalizing */
  if ((_desc(a).class == SUBNORMAL) && (_desc(b).class == SUBNORMAL))
  {
    sc_val_from_ulong(1, NULL);
    _shift_left(_mant(result), sc_get_buffer(), _mant(result));
  }

  /* resulting exponent is the bigger one */
  memmove(_exp(result), _exp(a), VALUE_SIZE);

  return _normalize(result, result, sticky);
541
542
}

543
static char* _mul(const char* a, const char* b, char* result)
544
{
545
  char *temp;
546

547
548
549
  if (_desc(a).class == NAN) {
    if (a != result) memcpy(result, a, CALC_BUFFER_SIZE);
    return result;
550
  }
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
  if (_desc(b).class == NAN) {
    if (b != result) memcpy(result, b, CALC_BUFFER_SIZE);
    return result;
  }

  temp = alloca(VALUE_SIZE);

  if (result != a && result != b)
    memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t));

  _sign(result) = _sign(a) ^ _sign(b);

  /* produce nan on 0 * inf */
  if (_desc(a).class == ZERO) {
    if (_desc(b).class == INF)
      fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    else
      if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }
  if (_desc(b).class == ZERO) {
    if (_desc(a).class == INF)
      fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    else
      if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }

  if (_desc(a).class == INF) {
    if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }
  if (_desc(b).class == INF) {
    if (b != result) memcpy(result+SIGN_POS+1, b+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }

  /* exp = exp(a) + exp(b) - excess */
  sc_add(_exp(a), _exp(b), _exp(result));

  sc_val_from_ulong((1<<_desc(a).exponent_size)/2-1, temp);
  sc_sub(_exp(result), temp, _exp(result));

  /* mixed normal, subnormal values introduce an error of 1, correct it */
  if ((_desc(a).class == SUBNORMAL) ^ (_desc(b).class == SUBNORMAL))
  {
    sc_val_from_ulong(1, temp);
    sc_add(_exp(result), temp, _exp(result));
  }

  sc_mul(_mant(a), _mant(b), _mant(result));

  /* realign result: after a multiplication the digits right of the radix
   * point are the sum of the factors' digits after the radix point. As all
   * values are normalized they both have the same amount of these digits,
   * which has to be restored by proper shifting
   * +2 because of the two rounding bits */
  sc_val_from_ulong(2 + _desc(result).mantissa_size, temp);

  _shift_right(_mant(result), temp, _mant(result));

  return _normalize(result, result, sc_had_carry());
613
614
}

615
static char* _div(const char* a, const char* b, char* result)
616
{
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
  char *temp, *dividend;

  if (_desc(a).class == NAN) {
    if (a != result) memcpy(result, a, CALC_BUFFER_SIZE);
    return result;
  }
  if (_desc(b).class == NAN) {
    if (b != result) memcpy(result, b, CALC_BUFFER_SIZE);
    return result;
  }

  temp = alloca(VALUE_SIZE);
  dividend = alloca(VALUE_SIZE);

  if (result != a && result != b)
    memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t));

  _sign(result) = _sign(a) ^ _sign(b);

  /* produce nan on 0/0 and inf/inf */
  if (_desc(a).class == ZERO) {
    if (_desc(b).class == ZERO)
      /* 0/0 -> nan */
      fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    else
      /* 0/x -> a */
      if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }

  if (_desc(b).class == INF) {
    if (_desc(a).class == INF)
      /* inf/inf -> nan */
      fc_get_qnan(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    else {
      /* x/inf -> 0 */
      sc_val_from_ulong(0, NULL);
      _save_result(_exp(result));
      _save_result(_mant(result));
      _desc(result).class = ZERO;
    }
    return result;
  }

  if (_desc(a).class == INF) {
    /* inf/x -> inf */
    if (a != result) memcpy(result+SIGN_POS+1, a+SIGN_POS+1, CALC_BUFFER_SIZE-1);
    return result;
  }
  if (_desc(b).class == ZERO) {
    /* division by zero */
    if (_sign(result))
      fc_get_minusinf(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    else
      fc_get_plusinf(_desc(a).exponent_size, _desc(a).mantissa_size, result);
    return result;
  }

  /* exp = exp(a) - exp(b) + excess - 1*/
  sc_sub(_exp(a), _exp(b), _exp(result));
  sc_val_from_ulong((1 << _desc(a).exponent_size)/2-2, temp);
  sc_add(_exp(result), temp, _exp(result));

  /* mixed normal, subnormal values introduce an error of 1, correct it */
  if ((_desc(a).class == SUBNORMAL) ^ (_desc(b).class == SUBNORMAL))
682
  {
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
    sc_val_from_ulong(1, temp);
    sc_add(_exp(result), temp, _exp(result));
  }

  /* mant(res) = mant(a) / 1/2mant(b) */
  /* to gain more bits of precision in the result the dividend could be
   * shifted left, as this operation does not loose bits. This would not
   * fit into the integer precision, but due to the rounding bits (which
   * are always zero because the values are all normalized) the divisor
   * can be shifted right instead to achieve the same result */
  sc_val_from_ulong(2 + _desc(result).mantissa_size, temp);

  _shift_left(_mant(a), temp, dividend);

  {
    char *divisor = alloca(CALC_BUFFER_SIZE);
    sc_val_from_ulong(1, divisor);
    _shift_right(_mant(b), divisor, divisor);
    sc_div(dividend, divisor, _mant(result));
702
  }
703
704

  return _normalize(result, result, sc_had_carry());
705
706
}

707
void _power_of_ten(int exp, descriptor_t *desc, char *result)
708
{
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
  char *build;
  char *temp;

  /* positive sign */
  _sign(result) = 0;

  /* set new descriptor (else result is supposed to already have one) */
  if (desc != NULL)
    memcpy(&_desc(result), desc, sizeof(descriptor_t));

  build = alloca(VALUE_SIZE);
  temp = alloca(VALUE_SIZE);

  sc_val_from_ulong((1 << _desc(result).exponent_size)/2-1, _exp(result));

  if (exp > 0)
  {
    /* temp is value of ten now */
    sc_val_from_ulong(10, NULL);
    _save_result(temp);

    for (exp--; exp > 0; exp--) {
      _save_result(build);
      sc_mul(build, temp, NULL);
    }
    _save_result(build);

    /* temp is amount of leftshift needed to put the value left of the radix point */
    sc_val_from_ulong(_desc(result).mantissa_size + 2, temp);
738

739
740
741
742
    _shift_left(build, temp, _mant(result));

    _normalize(result, result, 0);
  }
743
744
}

745
static char* _trunc(const char *a, char *result)
746
{
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
  /* when exponent == 0 all bits left of the radix point
   * are the integral part of the value. For 15bit exp_size
   * this would require a leftshift of max. 16383 bits which
   * is too much.
   * But it is enough to ensure that no bit right of the radix
   * point remains set. This restricts the interesting
   * exponents to the interval [0, mant_size-1].
   * Outside this interval the truncated value is either 0 or
   * it is are already truncated */

  int exp_bias, exp_val;
  char *temp;

  temp = alloca(VALUE_SIZE);

  if (a != result)
    memcpy(&_desc(result), &_desc(a), sizeof(descriptor_t));

  exp_bias = (1<<_desc(a).exponent_size)/2-1;
  exp_val = sc_val_to_long(_exp(a)) - exp_bias;

  if (exp_val < 0) {
    sc_val_from_ulong(0, NULL);
    _save_result(_exp(result));
    _save_result(_mant(result));
    _desc(result).class = ZERO;

    return result;
  }

  if (exp_val > _desc(a).mantissa_size) {
    if (a != result)
      memcpy(result, a, CALC_BUFFER_SIZE);

    return result;
  }

  /* set up a proper mask to delete all bits right of the
   * radix point if the mantissa had been shifted until exp == 0 */
  sc_max_from_bits(1 + exp_val, 0, temp);
  sc_val_from_long(_desc(a).mantissa_size - exp_val + 2, NULL);
  _shift_left(temp, sc_get_buffer(), temp);

  /* and the mask and return the result */
  sc_and(_mant(a), temp, _mant(result));

  if (a != result) memcpy(_exp(result), _exp(a), VALUE_SIZE);

  return result;
796
797
}

798
799
800
801
802
/*
 * This does value sanity checking(or should do it), sets up any prerequisites,
 * calls the proper internal functions, clears up and returns
 * the result */
char* _calc(const char *a, const char *b, int opcode, char *result)
803
{
804
805
806
807
808
809
810
811
812
  char *temp;
#ifdef FLTCALC_TRACE_CALC
  char *buffer;

  buffer = alloca(100);
#endif

  if (result == NULL) result = calc_buffer;

813
  TRACEPRINTF(("%s ", fc_print(a, buffer, 100, FC_PACKED)));
814
815
  switch (opcode)
  {
816
817
    case FC_add:
      /* make the value with the bigger exponent the first one */
818
      TRACEPRINTF(("+ %s ", fc_print(b, buffer, 100, FC_PACKED)));
819
820
821
822
823
824
      if (sc_comp(_exp(a), _exp(b)) == -1)
        _add(b, a, result);
      else
        _add(a, b, result);
      break;
    case FC_sub:
825
      TRACEPRINTF(("- %s ", fc_print(b, buffer, 100, FC_PACKED)));
826
827
828
829
830
831
832
      temp = alloca(CALC_BUFFER_SIZE);
      memcpy(temp, b, CALC_BUFFER_SIZE);
      _sign(temp) = !_sign(b);
      if (sc_comp(_exp(a), _exp(temp)) == -1)
        _add(temp, a, result);
      else
        _add(a, temp, result);
833
      break;
834
    case FC_mul:
835
      TRACEPRINTF(("* %s ", fc_print(b, buffer, 100, FC_PACKED)));
836
      _mul(a, b, result);
837
      break;
838
    case FC_div:
839
      TRACEPRINTF(("/ %s ", fc_print(b, buffer, 100, FC_PACKED)));
840
      _div(a, b, result);
841
      break;
842
843
844
845
    case FC_neg:
      TRACEPRINTF(("negated "));
      if (a != result) memcpy(result, a, CALC_BUFFER_SIZE);
      _sign(result) = !_sign(a);
846
      break;
847
848
849
850
    case FC_int:
      _trunc(a, result);
      break;
    case FC_rnd:
851
      break;
852
  }
853

854
  TRACEPRINTF(("= %s\n", fc_print(result, buffer, 100, FC_PACKED)));
855
  return result;
856
857
}

858
859
860
861
/********
 * functions defined in fltcalc.h
 ********/
const void *fc_get_buffer(void)
862
{
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
  return calc_buffer;
}

const int fc_get_buffer_length(void)
{
  return CALC_BUFFER_SIZE;
}

char* fc_val_from_str(const char *str, unsigned int len, char exp_size, char mant_size, char *result)
{
#if 0
  enum {
    START,
    LEFT_OF_DOT,
    RIGHT_OF_DOT,
    EXP_START,
    EXPONENT,
    END
  };

  char exp_sign;
  int exp_int, hsb, state;

  const char *old_str;

  int pos;
  char *mant_str, *exp_val, *power_val;

  if (result == NULL) result = calc_buffer;

  exp_val = alloca(VALUE_SIZE);
  power_val = alloca(CALC_BUFFER_SIZE);
  mant_str = alloca((len)?(len):(strlen(str)));

  _desc(result).exponent_size = exp_size;
  _desc(result).mantissa_size = mant_size;
  _desc(result).class = NORMAL;

  old_str = str;
  pos = 0;
  exp_int = 0;
  state = START;

  while (len == 0 || str-old_str < len)
  {
    switch (state) {
      case START:
        switch (*str) {
          case '+':
            _sign(result) = 0;
            state = LEFT_OF_DOT;
            str++;
            break;

          case '-':
            _sign(result) = 1;
            state = LEFT_OF_DOT;
            str++;
            break;

          case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
            _sign(result) = 0;
            state = LEFT_OF_DOT;
            break;

          case '.':
            _sign(result) = 0;
            state = RIGHT_OF_DOT;
            str++;
            break;

          case 'n':
          case 'N':
          case 'i':
          case 'I':
            break;

          default:
            _fail_char(old_str, len, str - old_str);
        }
        break;

      case LEFT_OF_DOT:
        switch (*str) {
          case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
            mant_str[pos++] = *(str++);
            break;

          case '.':
            state = RIGHT_OF_DOT;
            str++;
            break;

          case 'e':
          case 'E':
            state = EXP_START;
            str++;
            break;

          case '\0':
            mant_str[pos] = '\0';
            goto done;

          default:
            _fail_char(old_str, len, str - old_str);
        }
        break;

      case RIGHT_OF_DOT:
        switch (*str) {
          case '0': case '1': case '2': case '3': case '4': case '5': case '6': case '7': case '8': case '9':
            mant_str[pos++] = *(str++);
            exp_int++;
            break;

          case 'e':
          case 'E':
            state = EXP_START;
            str++;
            break;

          case '\0':
            mant_str[pos] = '\0';
            goto done;

          default:
            _fail_char(old_str, len, str - old_str);
        }
        break;

      case EXP_START:
        switch (*str) {
          case '-':
            exp_sign = 1;
            /* fall through */
          case '+':
            if (*(str-1) != 'e' && *(str-1) != 'E') _fail_char(old_str, len, str - old_str);
            str++;