2013-11-27 Kugan Vivekanandarajah <kuganv@linaro.org>
[official-gcc.git] / libgcc / fp-bit.c
blob9b0c194604f81d82492da1e615c1b8b99d25af6f
1 /* This is a software floating point library which can be used
2 for targets without hardware floating point.
3 Copyright (C) 1994-2013 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 Under Section 7 of GPL version 3, you are granted additional
18 permissions described in the GCC Runtime Library Exception, version
19 3.1, as published by the Free Software Foundation.
21 You should have received a copy of the GNU General Public License and
22 a copy of the GCC Runtime Library Exception along with this program;
23 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
24 <http://www.gnu.org/licenses/>. */
26 /* This implements IEEE 754 format arithmetic, but does not provide a
27 mechanism for setting the rounding mode, or for generating or handling
28 exceptions.
30 The original code by Steve Chamberlain, hacked by Mark Eichin and Jim
31 Wilson, all of Cygnus Support. */
33 /* The intended way to use this file is to make two copies, add `#define FLOAT'
34 to one copy, then compile both copies and add them to libgcc.a. */
36 #include "tconfig.h"
37 #include "coretypes.h"
38 #include "tm.h"
39 #include "libgcc_tm.h"
40 #include "fp-bit.h"
42 /* The following macros can be defined to change the behavior of this file:
43 FLOAT: Implement a `float', aka SFmode, fp library. If this is not
44 defined, then this file implements a `double', aka DFmode, fp library.
45 FLOAT_ONLY: Used with FLOAT, to implement a `float' only library, i.e.
46 don't include float->double conversion which requires the double library.
47 This is useful only for machines which can't support doubles, e.g. some
48 8-bit processors.
49 CMPtype: Specify the type that floating point compares should return.
50 This defaults to SItype, aka int.
51 _DEBUG_BITFLOAT: This makes debugging the code a little easier, by adding
52 two integers to the FLO_union_type.
53 NO_DENORMALS: Disable handling of denormals.
54 NO_NANS: Disable nan and infinity handling
55 SMALL_MACHINE: Useful when operations on QIs and HIs are faster
56 than on an SI */
58 /* We don't currently support extended floats (long doubles) on machines
59 without hardware to deal with them.
61 These stubs are just to keep the linker from complaining about unresolved
62 references which can be pulled in from libio & libstdc++, even if the
63 user isn't using long doubles. However, they may generate an unresolved
64 external to abort if abort is not used by the function, and the stubs
65 are referenced from within libc, since libgcc goes before and after the
66 system library. */
68 #ifdef DECLARE_LIBRARY_RENAMES
69 DECLARE_LIBRARY_RENAMES
70 #endif
72 #ifdef EXTENDED_FLOAT_STUBS
73 extern void abort (void);
74 void __extendsfxf2 (void) { abort(); }
75 void __extenddfxf2 (void) { abort(); }
76 void __truncxfdf2 (void) { abort(); }
77 void __truncxfsf2 (void) { abort(); }
78 void __fixxfsi (void) { abort(); }
79 void __floatsixf (void) { abort(); }
80 void __addxf3 (void) { abort(); }
81 void __subxf3 (void) { abort(); }
82 void __mulxf3 (void) { abort(); }
83 void __divxf3 (void) { abort(); }
84 void __negxf2 (void) { abort(); }
85 void __eqxf2 (void) { abort(); }
86 void __nexf2 (void) { abort(); }
87 void __gtxf2 (void) { abort(); }
88 void __gexf2 (void) { abort(); }
89 void __lexf2 (void) { abort(); }
90 void __ltxf2 (void) { abort(); }
92 void __extendsftf2 (void) { abort(); }
93 void __extenddftf2 (void) { abort(); }
94 void __trunctfdf2 (void) { abort(); }
95 void __trunctfsf2 (void) { abort(); }
96 void __fixtfsi (void) { abort(); }
97 void __floatsitf (void) { abort(); }
98 void __addtf3 (void) { abort(); }
99 void __subtf3 (void) { abort(); }
100 void __multf3 (void) { abort(); }
101 void __divtf3 (void) { abort(); }
102 void __negtf2 (void) { abort(); }
103 void __eqtf2 (void) { abort(); }
104 void __netf2 (void) { abort(); }
105 void __gttf2 (void) { abort(); }
106 void __getf2 (void) { abort(); }
107 void __letf2 (void) { abort(); }
108 void __lttf2 (void) { abort(); }
109 #else /* !EXTENDED_FLOAT_STUBS, rest of file */
111 /* IEEE "special" number predicates */
113 #ifdef NO_NANS
115 #define nan() 0
116 #define isnan(x) 0
117 #define isinf(x) 0
118 #else
120 #if defined L_thenan_sf
121 const fp_number_type __thenan_sf = { CLASS_SNAN, 0, 0, {(fractype) 0} };
122 #elif defined L_thenan_df
123 const fp_number_type __thenan_df = { CLASS_SNAN, 0, 0, {(fractype) 0} };
124 #elif defined L_thenan_tf
125 const fp_number_type __thenan_tf = { CLASS_SNAN, 0, 0, {(fractype) 0} };
126 #elif defined TFLOAT
127 extern const fp_number_type __thenan_tf;
128 #elif defined FLOAT
129 extern const fp_number_type __thenan_sf;
130 #else
131 extern const fp_number_type __thenan_df;
132 #endif
134 INLINE
135 static const fp_number_type *
136 makenan (void)
138 #ifdef TFLOAT
139 return & __thenan_tf;
140 #elif defined FLOAT
141 return & __thenan_sf;
142 #else
143 return & __thenan_df;
144 #endif
147 INLINE
148 static int
149 isnan (const fp_number_type *x)
151 return __builtin_expect (x->class == CLASS_SNAN || x->class == CLASS_QNAN,
155 INLINE
156 static int
157 isinf (const fp_number_type * x)
159 return __builtin_expect (x->class == CLASS_INFINITY, 0);
162 #endif /* NO_NANS */
164 INLINE
165 static int
166 iszero (const fp_number_type * x)
168 return x->class == CLASS_ZERO;
171 INLINE
172 static void
173 flip_sign ( fp_number_type * x)
175 x->sign = !x->sign;
178 /* Count leading zeroes in N. */
179 INLINE
180 static int
181 clzusi (USItype n)
183 extern int __clzsi2 (USItype);
184 if (sizeof (USItype) == sizeof (unsigned int))
185 return __builtin_clz (n);
186 else if (sizeof (USItype) == sizeof (unsigned long))
187 return __builtin_clzl (n);
188 else if (sizeof (USItype) == sizeof (unsigned long long))
189 return __builtin_clzll (n);
190 else
191 return __clzsi2 (n);
194 extern FLO_type pack_d (const fp_number_type * );
196 #if defined(L_pack_df) || defined(L_pack_sf) || defined(L_pack_tf)
197 FLO_type
198 pack_d (const fp_number_type *src)
200 FLO_union_type dst;
201 fractype fraction = src->fraction.ll; /* wasn't unsigned before? */
202 int sign = src->sign;
203 int exp = 0;
205 if (LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS) && (isnan (src) || isinf (src)))
207 /* We can't represent these values accurately. By using the
208 largest possible magnitude, we guarantee that the conversion
209 of infinity is at least as big as any finite number. */
210 exp = EXPMAX;
211 fraction = ((fractype) 1 << FRACBITS) - 1;
213 else if (isnan (src))
215 exp = EXPMAX;
216 /* Restore the NaN's payload. */
217 fraction >>= NGARDS;
218 fraction &= QUIET_NAN - 1;
219 if (src->class == CLASS_QNAN || 1)
221 #ifdef QUIET_NAN_NEGATED
222 /* The quiet/signaling bit remains unset. */
223 /* Make sure the fraction has a non-zero value. */
224 if (fraction == 0)
225 fraction |= QUIET_NAN - 1;
226 #else
227 /* Set the quiet/signaling bit. */
228 fraction |= QUIET_NAN;
229 #endif
232 else if (isinf (src))
234 exp = EXPMAX;
235 fraction = 0;
237 else if (iszero (src))
239 exp = 0;
240 fraction = 0;
242 else if (fraction == 0)
244 exp = 0;
246 else
248 if (__builtin_expect (src->normal_exp < NORMAL_EXPMIN, 0))
250 #ifdef NO_DENORMALS
251 /* Go straight to a zero representation if denormals are not
252 supported. The denormal handling would be harmless but
253 isn't unnecessary. */
254 exp = 0;
255 fraction = 0;
256 #else /* NO_DENORMALS */
257 /* This number's exponent is too low to fit into the bits
258 available in the number, so we'll store 0 in the exponent and
259 shift the fraction to the right to make up for it. */
261 int shift = NORMAL_EXPMIN - src->normal_exp;
263 exp = 0;
265 if (shift > FRAC_NBITS - NGARDS)
267 /* No point shifting, since it's more that 64 out. */
268 fraction = 0;
270 else
272 int lowbit = (fraction & (((fractype)1 << shift) - 1)) ? 1 : 0;
273 fraction = (fraction >> shift) | lowbit;
275 if ((fraction & GARDMASK) == GARDMSB)
277 if ((fraction & (1 << NGARDS)))
278 fraction += GARDROUND + 1;
280 else
282 /* Add to the guards to round up. */
283 fraction += GARDROUND;
285 /* Perhaps the rounding means we now need to change the
286 exponent, because the fraction is no longer denormal. */
287 if (fraction >= IMPLICIT_1)
289 exp += 1;
291 fraction >>= NGARDS;
292 #endif /* NO_DENORMALS */
294 else if (!LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS)
295 && __builtin_expect (src->normal_exp > EXPBIAS, 0))
297 exp = EXPMAX;
298 fraction = 0;
300 else
302 exp = src->normal_exp + EXPBIAS;
303 if (!ROUND_TOWARDS_ZERO)
305 /* IF the gard bits are the all zero, but the first, then we're
306 half way between two numbers, choose the one which makes the
307 lsb of the answer 0. */
308 if ((fraction & GARDMASK) == GARDMSB)
310 if (fraction & (1 << NGARDS))
311 fraction += GARDROUND + 1;
313 else
315 /* Add a one to the guards to round up */
316 fraction += GARDROUND;
318 if (fraction >= IMPLICIT_2)
320 fraction >>= 1;
321 exp += 1;
324 fraction >>= NGARDS;
326 if (LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS) && exp > EXPMAX)
328 /* Saturate on overflow. */
329 exp = EXPMAX;
330 fraction = ((fractype) 1 << FRACBITS) - 1;
335 /* We previously used bitfields to store the number, but this doesn't
336 handle little/big endian systems conveniently, so use shifts and
337 masks */
338 #ifdef FLOAT_BIT_ORDER_MISMATCH
339 dst.bits.fraction = fraction;
340 dst.bits.exp = exp;
341 dst.bits.sign = sign;
342 #else
343 # if defined TFLOAT && defined HALFFRACBITS
345 halffractype high, low, unity;
346 int lowsign, lowexp;
348 unity = (halffractype) 1 << HALFFRACBITS;
350 /* Set HIGH to the high double's significand, masking out the implicit 1.
351 Set LOW to the low double's full significand. */
352 high = (fraction >> (FRACBITS - HALFFRACBITS)) & (unity - 1);
353 low = fraction & (unity * 2 - 1);
355 /* Get the initial sign and exponent of the low double. */
356 lowexp = exp - HALFFRACBITS - 1;
357 lowsign = sign;
359 /* HIGH should be rounded like a normal double, making |LOW| <=
360 0.5 ULP of HIGH. Assume round-to-nearest. */
361 if (exp < EXPMAX)
362 if (low > unity || (low == unity && (high & 1) == 1))
364 /* Round HIGH up and adjust LOW to match. */
365 high++;
366 if (high == unity)
368 /* May make it infinite, but that's OK. */
369 high = 0;
370 exp++;
372 low = unity * 2 - low;
373 lowsign ^= 1;
376 high |= (halffractype) exp << HALFFRACBITS;
377 high |= (halffractype) sign << (HALFFRACBITS + EXPBITS);
379 if (exp == EXPMAX || exp == 0 || low == 0)
380 low = 0;
381 else
383 while (lowexp > 0 && low < unity)
385 low <<= 1;
386 lowexp--;
389 if (lowexp <= 0)
391 halffractype roundmsb, round;
392 int shift;
394 shift = 1 - lowexp;
395 roundmsb = (1 << (shift - 1));
396 round = low & ((roundmsb << 1) - 1);
398 low >>= shift;
399 lowexp = 0;
401 if (round > roundmsb || (round == roundmsb && (low & 1) == 1))
403 low++;
404 if (low == unity)
405 /* LOW rounds up to the smallest normal number. */
406 lowexp++;
410 low &= unity - 1;
411 low |= (halffractype) lowexp << HALFFRACBITS;
412 low |= (halffractype) lowsign << (HALFFRACBITS + EXPBITS);
414 dst.value_raw = ((fractype) high << HALFSHIFT) | low;
416 # else
417 dst.value_raw = fraction & ((((fractype)1) << FRACBITS) - (fractype)1);
418 dst.value_raw |= ((fractype) (exp & ((1 << EXPBITS) - 1))) << FRACBITS;
419 dst.value_raw |= ((fractype) (sign & 1)) << (FRACBITS | EXPBITS);
420 # endif
421 #endif
423 #if defined(FLOAT_WORD_ORDER_MISMATCH) && !defined(FLOAT)
424 #ifdef TFLOAT
426 qrtrfractype tmp1 = dst.words[0];
427 qrtrfractype tmp2 = dst.words[1];
428 dst.words[0] = dst.words[3];
429 dst.words[1] = dst.words[2];
430 dst.words[2] = tmp2;
431 dst.words[3] = tmp1;
433 #else
435 halffractype tmp = dst.words[0];
436 dst.words[0] = dst.words[1];
437 dst.words[1] = tmp;
439 #endif
440 #endif
442 return dst.value;
444 #endif
446 #if defined(L_unpack_df) || defined(L_unpack_sf) || defined(L_unpack_tf)
447 void
448 unpack_d (FLO_union_type * src, fp_number_type * dst)
450 /* We previously used bitfields to store the number, but this doesn't
451 handle little/big endian systems conveniently, so use shifts and
452 masks */
453 fractype fraction;
454 int exp;
455 int sign;
457 #if defined(FLOAT_WORD_ORDER_MISMATCH) && !defined(FLOAT)
458 FLO_union_type swapped;
460 #ifdef TFLOAT
461 swapped.words[0] = src->words[3];
462 swapped.words[1] = src->words[2];
463 swapped.words[2] = src->words[1];
464 swapped.words[3] = src->words[0];
465 #else
466 swapped.words[0] = src->words[1];
467 swapped.words[1] = src->words[0];
468 #endif
469 src = &swapped;
470 #endif
472 #ifdef FLOAT_BIT_ORDER_MISMATCH
473 fraction = src->bits.fraction;
474 exp = src->bits.exp;
475 sign = src->bits.sign;
476 #else
477 # if defined TFLOAT && defined HALFFRACBITS
479 halffractype high, low;
481 high = src->value_raw >> HALFSHIFT;
482 low = src->value_raw & (((fractype)1 << HALFSHIFT) - 1);
484 fraction = high & ((((fractype)1) << HALFFRACBITS) - 1);
485 fraction <<= FRACBITS - HALFFRACBITS;
486 exp = ((int)(high >> HALFFRACBITS)) & ((1 << EXPBITS) - 1);
487 sign = ((int)(high >> (((HALFFRACBITS + EXPBITS))))) & 1;
489 if (exp != EXPMAX && exp != 0 && low != 0)
491 int lowexp = ((int)(low >> HALFFRACBITS)) & ((1 << EXPBITS) - 1);
492 int lowsign = ((int)(low >> (((HALFFRACBITS + EXPBITS))))) & 1;
493 int shift;
494 fractype xlow;
496 xlow = low & ((((fractype)1) << HALFFRACBITS) - 1);
497 if (lowexp)
498 xlow |= (((halffractype)1) << HALFFRACBITS);
499 else
500 lowexp = 1;
501 shift = (FRACBITS - HALFFRACBITS) - (exp - lowexp);
502 if (shift > 0)
503 xlow <<= shift;
504 else if (shift < 0)
505 xlow >>= -shift;
506 if (sign == lowsign)
507 fraction += xlow;
508 else if (fraction >= xlow)
509 fraction -= xlow;
510 else
512 /* The high part is a power of two but the full number is lower.
513 This code will leave the implicit 1 in FRACTION, but we'd
514 have added that below anyway. */
515 fraction = (((fractype) 1 << FRACBITS) - xlow) << 1;
516 exp--;
520 # else
521 fraction = src->value_raw & ((((fractype)1) << FRACBITS) - 1);
522 exp = ((int)(src->value_raw >> FRACBITS)) & ((1 << EXPBITS) - 1);
523 sign = ((int)(src->value_raw >> (FRACBITS + EXPBITS))) & 1;
524 # endif
525 #endif
527 dst->sign = sign;
528 if (exp == 0)
530 /* Hmm. Looks like 0 */
531 if (fraction == 0
532 #ifdef NO_DENORMALS
533 || 1
534 #endif
537 /* tastes like zero */
538 dst->class = CLASS_ZERO;
540 else
542 /* Zero exponent with nonzero fraction - it's denormalized,
543 so there isn't a leading implicit one - we'll shift it so
544 it gets one. */
545 dst->normal_exp = exp - EXPBIAS + 1;
546 fraction <<= NGARDS;
548 dst->class = CLASS_NUMBER;
549 #if 1
550 while (fraction < IMPLICIT_1)
552 fraction <<= 1;
553 dst->normal_exp--;
555 #endif
556 dst->fraction.ll = fraction;
559 else if (!LARGEST_EXPONENT_IS_NORMAL (FRAC_NBITS)
560 && __builtin_expect (exp == EXPMAX, 0))
562 /* Huge exponent*/
563 if (fraction == 0)
565 /* Attached to a zero fraction - means infinity */
566 dst->class = CLASS_INFINITY;
568 else
570 /* Nonzero fraction, means nan */
571 #ifdef QUIET_NAN_NEGATED
572 if ((fraction & QUIET_NAN) == 0)
573 #else
574 if (fraction & QUIET_NAN)
575 #endif
577 dst->class = CLASS_QNAN;
579 else
581 dst->class = CLASS_SNAN;
583 /* Now that we know which kind of NaN we got, discard the
584 quiet/signaling bit, but do preserve the NaN payload. */
585 fraction &= ~QUIET_NAN;
586 dst->fraction.ll = fraction << NGARDS;
589 else
591 /* Nothing strange about this number */
592 dst->normal_exp = exp - EXPBIAS;
593 dst->class = CLASS_NUMBER;
594 dst->fraction.ll = (fraction << NGARDS) | IMPLICIT_1;
597 #endif /* L_unpack_df || L_unpack_sf */
599 #if defined(L_addsub_sf) || defined(L_addsub_df) || defined(L_addsub_tf)
600 static const fp_number_type *
601 _fpadd_parts (fp_number_type * a,
602 fp_number_type * b,
603 fp_number_type * tmp)
605 intfrac tfraction;
607 /* Put commonly used fields in local variables. */
608 int a_normal_exp;
609 int b_normal_exp;
610 fractype a_fraction;
611 fractype b_fraction;
613 if (isnan (a))
615 return a;
617 if (isnan (b))
619 return b;
621 if (isinf (a))
623 /* Adding infinities with opposite signs yields a NaN. */
624 if (isinf (b) && a->sign != b->sign)
625 return makenan ();
626 return a;
628 if (isinf (b))
630 return b;
632 if (iszero (b))
634 if (iszero (a))
636 *tmp = *a;
637 tmp->sign = a->sign & b->sign;
638 return tmp;
640 return a;
642 if (iszero (a))
644 return b;
647 /* Got two numbers. shift the smaller and increment the exponent till
648 they're the same */
650 int diff;
651 int sdiff;
653 a_normal_exp = a->normal_exp;
654 b_normal_exp = b->normal_exp;
655 a_fraction = a->fraction.ll;
656 b_fraction = b->fraction.ll;
658 diff = a_normal_exp - b_normal_exp;
659 sdiff = diff;
661 if (diff < 0)
662 diff = -diff;
663 if (diff < FRAC_NBITS)
665 if (sdiff > 0)
667 b_normal_exp += diff;
668 LSHIFT (b_fraction, diff);
670 else if (sdiff < 0)
672 a_normal_exp += diff;
673 LSHIFT (a_fraction, diff);
676 else
678 /* Somethings's up.. choose the biggest */
679 if (a_normal_exp > b_normal_exp)
681 b_normal_exp = a_normal_exp;
682 b_fraction = 0;
684 else
686 a_normal_exp = b_normal_exp;
687 a_fraction = 0;
692 if (a->sign != b->sign)
694 if (a->sign)
696 tfraction = -a_fraction + b_fraction;
698 else
700 tfraction = a_fraction - b_fraction;
702 if (tfraction >= 0)
704 tmp->sign = 0;
705 tmp->normal_exp = a_normal_exp;
706 tmp->fraction.ll = tfraction;
708 else
710 tmp->sign = 1;
711 tmp->normal_exp = a_normal_exp;
712 tmp->fraction.ll = -tfraction;
714 /* and renormalize it */
716 while (tmp->fraction.ll < IMPLICIT_1 && tmp->fraction.ll)
718 tmp->fraction.ll <<= 1;
719 tmp->normal_exp--;
722 else
724 tmp->sign = a->sign;
725 tmp->normal_exp = a_normal_exp;
726 tmp->fraction.ll = a_fraction + b_fraction;
728 tmp->class = CLASS_NUMBER;
729 /* Now the fraction is added, we have to shift down to renormalize the
730 number */
732 if (tmp->fraction.ll >= IMPLICIT_2)
734 LSHIFT (tmp->fraction.ll, 1);
735 tmp->normal_exp++;
737 return tmp;
740 FLO_type
741 add (FLO_type arg_a, FLO_type arg_b)
743 fp_number_type a;
744 fp_number_type b;
745 fp_number_type tmp;
746 const fp_number_type *res;
747 FLO_union_type au, bu;
749 au.value = arg_a;
750 bu.value = arg_b;
752 unpack_d (&au, &a);
753 unpack_d (&bu, &b);
755 res = _fpadd_parts (&a, &b, &tmp);
757 return pack_d (res);
760 FLO_type
761 sub (FLO_type arg_a, FLO_type arg_b)
763 fp_number_type a;
764 fp_number_type b;
765 fp_number_type tmp;
766 const fp_number_type *res;
767 FLO_union_type au, bu;
769 au.value = arg_a;
770 bu.value = arg_b;
772 unpack_d (&au, &a);
773 unpack_d (&bu, &b);
775 b.sign ^= 1;
777 res = _fpadd_parts (&a, &b, &tmp);
779 return pack_d (res);
781 #endif /* L_addsub_sf || L_addsub_df */
783 #if defined(L_mul_sf) || defined(L_mul_df) || defined(L_mul_tf)
784 static inline __attribute__ ((__always_inline__)) const fp_number_type *
785 _fpmul_parts ( fp_number_type * a,
786 fp_number_type * b,
787 fp_number_type * tmp)
789 fractype low = 0;
790 fractype high = 0;
792 if (isnan (a))
794 a->sign = a->sign != b->sign;
795 return a;
797 if (isnan (b))
799 b->sign = a->sign != b->sign;
800 return b;
802 if (isinf (a))
804 if (iszero (b))
805 return makenan ();
806 a->sign = a->sign != b->sign;
807 return a;
809 if (isinf (b))
811 if (iszero (a))
813 return makenan ();
815 b->sign = a->sign != b->sign;
816 return b;
818 if (iszero (a))
820 a->sign = a->sign != b->sign;
821 return a;
823 if (iszero (b))
825 b->sign = a->sign != b->sign;
826 return b;
829 /* Calculate the mantissa by multiplying both numbers to get a
830 twice-as-wide number. */
832 #if defined(NO_DI_MODE) || defined(TFLOAT)
834 fractype x = a->fraction.ll;
835 fractype ylow = b->fraction.ll;
836 fractype yhigh = 0;
837 int bit;
839 /* ??? This does multiplies one bit at a time. Optimize. */
840 for (bit = 0; bit < FRAC_NBITS; bit++)
842 int carry;
844 if (x & 1)
846 carry = (low += ylow) < ylow;
847 high += yhigh + carry;
849 yhigh <<= 1;
850 if (ylow & FRACHIGH)
852 yhigh |= 1;
854 ylow <<= 1;
855 x >>= 1;
858 #elif defined(FLOAT)
859 /* Multiplying two USIs to get a UDI, we're safe. */
861 UDItype answer = (UDItype)a->fraction.ll * (UDItype)b->fraction.ll;
863 high = answer >> BITS_PER_SI;
864 low = answer;
866 #else
867 /* fractype is DImode, but we need the result to be twice as wide.
868 Assuming a widening multiply from DImode to TImode is not
869 available, build one by hand. */
871 USItype nl = a->fraction.ll;
872 USItype nh = a->fraction.ll >> BITS_PER_SI;
873 USItype ml = b->fraction.ll;
874 USItype mh = b->fraction.ll >> BITS_PER_SI;
875 UDItype pp_ll = (UDItype) ml * nl;
876 UDItype pp_hl = (UDItype) mh * nl;
877 UDItype pp_lh = (UDItype) ml * nh;
878 UDItype pp_hh = (UDItype) mh * nh;
879 UDItype res2 = 0;
880 UDItype res0 = 0;
881 UDItype ps_hh__ = pp_hl + pp_lh;
882 if (ps_hh__ < pp_hl)
883 res2 += (UDItype)1 << BITS_PER_SI;
884 pp_hl = (UDItype)(USItype)ps_hh__ << BITS_PER_SI;
885 res0 = pp_ll + pp_hl;
886 if (res0 < pp_ll)
887 res2++;
888 res2 += (ps_hh__ >> BITS_PER_SI) + pp_hh;
889 high = res2;
890 low = res0;
892 #endif
895 tmp->normal_exp = a->normal_exp + b->normal_exp
896 + FRAC_NBITS - (FRACBITS + NGARDS);
897 tmp->sign = a->sign != b->sign;
898 while (high >= IMPLICIT_2)
900 tmp->normal_exp++;
901 if (high & 1)
903 low >>= 1;
904 low |= FRACHIGH;
906 high >>= 1;
908 while (high < IMPLICIT_1)
910 tmp->normal_exp--;
912 high <<= 1;
913 if (low & FRACHIGH)
914 high |= 1;
915 low <<= 1;
918 if (!ROUND_TOWARDS_ZERO && (high & GARDMASK) == GARDMSB)
920 if (high & (1 << NGARDS))
922 /* Because we're half way, we would round to even by adding
923 GARDROUND + 1, except that's also done in the packing
924 function, and rounding twice will lose precision and cause
925 the result to be too far off. Example: 32-bit floats with
926 bit patterns 0xfff * 0x3f800400 ~= 0xfff (less than 0.5ulp
927 off), not 0x1000 (more than 0.5ulp off). */
929 else if (low)
931 /* We're a further than half way by a small amount corresponding
932 to the bits set in "low". Knowing that, we round here and
933 not in pack_d, because there we don't have "low" available
934 anymore. */
935 high += GARDROUND + 1;
937 /* Avoid further rounding in pack_d. */
938 high &= ~(fractype) GARDMASK;
941 tmp->fraction.ll = high;
942 tmp->class = CLASS_NUMBER;
943 return tmp;
946 FLO_type
947 multiply (FLO_type arg_a, FLO_type arg_b)
949 fp_number_type a;
950 fp_number_type b;
951 fp_number_type tmp;
952 const fp_number_type *res;
953 FLO_union_type au, bu;
955 au.value = arg_a;
956 bu.value = arg_b;
958 unpack_d (&au, &a);
959 unpack_d (&bu, &b);
961 res = _fpmul_parts (&a, &b, &tmp);
963 return pack_d (res);
965 #endif /* L_mul_sf || L_mul_df || L_mul_tf */
967 #if defined(L_div_sf) || defined(L_div_df) || defined(L_div_tf)
968 static inline __attribute__ ((__always_inline__)) const fp_number_type *
969 _fpdiv_parts (fp_number_type * a,
970 fp_number_type * b)
972 fractype bit;
973 fractype numerator;
974 fractype denominator;
975 fractype quotient;
977 if (isnan (a))
979 return a;
981 if (isnan (b))
983 return b;
986 a->sign = a->sign ^ b->sign;
988 if (isinf (a) || iszero (a))
990 if (a->class == b->class)
991 return makenan ();
992 return a;
995 if (isinf (b))
997 a->fraction.ll = 0;
998 a->normal_exp = 0;
999 return a;
1001 if (iszero (b))
1003 a->class = CLASS_INFINITY;
1004 return a;
1007 /* Calculate the mantissa by multiplying both 64bit numbers to get a
1008 128 bit number */
1010 /* quotient =
1011 ( numerator / denominator) * 2^(numerator exponent - denominator exponent)
1014 a->normal_exp = a->normal_exp - b->normal_exp;
1015 numerator = a->fraction.ll;
1016 denominator = b->fraction.ll;
1018 if (numerator < denominator)
1020 /* Fraction will be less than 1.0 */
1021 numerator *= 2;
1022 a->normal_exp--;
1024 bit = IMPLICIT_1;
1025 quotient = 0;
1026 /* ??? Does divide one bit at a time. Optimize. */
1027 while (bit)
1029 if (numerator >= denominator)
1031 quotient |= bit;
1032 numerator -= denominator;
1034 bit >>= 1;
1035 numerator *= 2;
1038 if (!ROUND_TOWARDS_ZERO && (quotient & GARDMASK) == GARDMSB)
1040 if (quotient & (1 << NGARDS))
1042 /* Because we're half way, we would round to even by adding
1043 GARDROUND + 1, except that's also done in the packing
1044 function, and rounding twice will lose precision and cause
1045 the result to be too far off. */
1047 else if (numerator)
1049 /* We're a further than half way by the small amount
1050 corresponding to the bits set in "numerator". Knowing
1051 that, we round here and not in pack_d, because there we
1052 don't have "numerator" available anymore. */
1053 quotient += GARDROUND + 1;
1055 /* Avoid further rounding in pack_d. */
1056 quotient &= ~(fractype) GARDMASK;
1060 a->fraction.ll = quotient;
1061 return (a);
1065 FLO_type
1066 divide (FLO_type arg_a, FLO_type arg_b)
1068 fp_number_type a;
1069 fp_number_type b;
1070 const fp_number_type *res;
1071 FLO_union_type au, bu;
1073 au.value = arg_a;
1074 bu.value = arg_b;
1076 unpack_d (&au, &a);
1077 unpack_d (&bu, &b);
1079 res = _fpdiv_parts (&a, &b);
1081 return pack_d (res);
1083 #endif /* L_div_sf || L_div_df */
1085 #if defined(L_fpcmp_parts_sf) || defined(L_fpcmp_parts_df) \
1086 || defined(L_fpcmp_parts_tf)
1087 /* according to the demo, fpcmp returns a comparison with 0... thus
1088 a<b -> -1
1089 a==b -> 0
1090 a>b -> +1
1094 __fpcmp_parts (fp_number_type * a, fp_number_type * b)
1096 #if 0
1097 /* either nan -> unordered. Must be checked outside of this routine. */
1098 if (isnan (a) && isnan (b))
1100 return 1; /* still unordered! */
1102 #endif
1104 if (isnan (a) || isnan (b))
1106 return 1; /* how to indicate unordered compare? */
1108 if (isinf (a) && isinf (b))
1110 /* +inf > -inf, but +inf != +inf */
1111 /* b \a| +inf(0)| -inf(1)
1112 ______\+--------+--------
1113 +inf(0)| a==b(0)| a<b(-1)
1114 -------+--------+--------
1115 -inf(1)| a>b(1) | a==b(0)
1116 -------+--------+--------
1117 So since unordered must be nonzero, just line up the columns...
1119 return b->sign - a->sign;
1121 /* but not both... */
1122 if (isinf (a))
1124 return a->sign ? -1 : 1;
1126 if (isinf (b))
1128 return b->sign ? 1 : -1;
1130 if (iszero (a) && iszero (b))
1132 return 0;
1134 if (iszero (a))
1136 return b->sign ? 1 : -1;
1138 if (iszero (b))
1140 return a->sign ? -1 : 1;
1142 /* now both are "normal". */
1143 if (a->sign != b->sign)
1145 /* opposite signs */
1146 return a->sign ? -1 : 1;
1148 /* same sign; exponents? */
1149 if (a->normal_exp > b->normal_exp)
1151 return a->sign ? -1 : 1;
1153 if (a->normal_exp < b->normal_exp)
1155 return a->sign ? 1 : -1;
1157 /* same exponents; check size. */
1158 if (a->fraction.ll > b->fraction.ll)
1160 return a->sign ? -1 : 1;
1162 if (a->fraction.ll < b->fraction.ll)
1164 return a->sign ? 1 : -1;
1166 /* after all that, they're equal. */
1167 return 0;
1169 #endif
1171 #if defined(L_compare_sf) || defined(L_compare_df) || defined(L_compoare_tf)
1172 CMPtype
1173 compare (FLO_type arg_a, FLO_type arg_b)
1175 fp_number_type a;
1176 fp_number_type b;
1177 FLO_union_type au, bu;
1179 au.value = arg_a;
1180 bu.value = arg_b;
1182 unpack_d (&au, &a);
1183 unpack_d (&bu, &b);
1185 return __fpcmp_parts (&a, &b);
1187 #endif /* L_compare_sf || L_compare_df */
1189 /* These should be optimized for their specific tasks someday. */
1191 #if defined(L_eq_sf) || defined(L_eq_df) || defined(L_eq_tf)
1192 CMPtype
1193 _eq_f2 (FLO_type arg_a, FLO_type arg_b)
1195 fp_number_type a;
1196 fp_number_type b;
1197 FLO_union_type au, bu;
1199 au.value = arg_a;
1200 bu.value = arg_b;
1202 unpack_d (&au, &a);
1203 unpack_d (&bu, &b);
1205 if (isnan (&a) || isnan (&b))
1206 return 1; /* false, truth == 0 */
1208 return __fpcmp_parts (&a, &b) ;
1210 #endif /* L_eq_sf || L_eq_df */
1212 #if defined(L_ne_sf) || defined(L_ne_df) || defined(L_ne_tf)
1213 CMPtype
1214 _ne_f2 (FLO_type arg_a, FLO_type arg_b)
1216 fp_number_type a;
1217 fp_number_type b;
1218 FLO_union_type au, bu;
1220 au.value = arg_a;
1221 bu.value = arg_b;
1223 unpack_d (&au, &a);
1224 unpack_d (&bu, &b);
1226 if (isnan (&a) || isnan (&b))
1227 return 1; /* true, truth != 0 */
1229 return __fpcmp_parts (&a, &b) ;
1231 #endif /* L_ne_sf || L_ne_df */
1233 #if defined(L_gt_sf) || defined(L_gt_df) || defined(L_gt_tf)
1234 CMPtype
1235 _gt_f2 (FLO_type arg_a, FLO_type arg_b)
1237 fp_number_type a;
1238 fp_number_type b;
1239 FLO_union_type au, bu;
1241 au.value = arg_a;
1242 bu.value = arg_b;
1244 unpack_d (&au, &a);
1245 unpack_d (&bu, &b);
1247 if (isnan (&a) || isnan (&b))
1248 return -1; /* false, truth > 0 */
1250 return __fpcmp_parts (&a, &b);
1252 #endif /* L_gt_sf || L_gt_df */
1254 #if defined(L_ge_sf) || defined(L_ge_df) || defined(L_ge_tf)
1255 CMPtype
1256 _ge_f2 (FLO_type arg_a, FLO_type arg_b)
1258 fp_number_type a;
1259 fp_number_type b;
1260 FLO_union_type au, bu;
1262 au.value = arg_a;
1263 bu.value = arg_b;
1265 unpack_d (&au, &a);
1266 unpack_d (&bu, &b);
1268 if (isnan (&a) || isnan (&b))
1269 return -1; /* false, truth >= 0 */
1270 return __fpcmp_parts (&a, &b) ;
1272 #endif /* L_ge_sf || L_ge_df */
1274 #if defined(L_lt_sf) || defined(L_lt_df) || defined(L_lt_tf)
1275 CMPtype
1276 _lt_f2 (FLO_type arg_a, FLO_type arg_b)
1278 fp_number_type a;
1279 fp_number_type b;
1280 FLO_union_type au, bu;
1282 au.value = arg_a;
1283 bu.value = arg_b;
1285 unpack_d (&au, &a);
1286 unpack_d (&bu, &b);
1288 if (isnan (&a) || isnan (&b))
1289 return 1; /* false, truth < 0 */
1291 return __fpcmp_parts (&a, &b);
1293 #endif /* L_lt_sf || L_lt_df */
1295 #if defined(L_le_sf) || defined(L_le_df) || defined(L_le_tf)
1296 CMPtype
1297 _le_f2 (FLO_type arg_a, FLO_type arg_b)
1299 fp_number_type a;
1300 fp_number_type b;
1301 FLO_union_type au, bu;
1303 au.value = arg_a;
1304 bu.value = arg_b;
1306 unpack_d (&au, &a);
1307 unpack_d (&bu, &b);
1309 if (isnan (&a) || isnan (&b))
1310 return 1; /* false, truth <= 0 */
1312 return __fpcmp_parts (&a, &b) ;
1314 #endif /* L_le_sf || L_le_df */
1316 #if defined(L_unord_sf) || defined(L_unord_df) || defined(L_unord_tf)
1317 CMPtype
1318 _unord_f2 (FLO_type arg_a, FLO_type arg_b)
1320 fp_number_type a;
1321 fp_number_type b;
1322 FLO_union_type au, bu;
1324 au.value = arg_a;
1325 bu.value = arg_b;
1327 unpack_d (&au, &a);
1328 unpack_d (&bu, &b);
1330 return (isnan (&a) || isnan (&b));
1332 #endif /* L_unord_sf || L_unord_df */
1334 #if defined(L_si_to_sf) || defined(L_si_to_df) || defined(L_si_to_tf)
1335 FLO_type
1336 si_to_float (SItype arg_a)
1338 fp_number_type in;
1340 in.class = CLASS_NUMBER;
1341 in.sign = arg_a < 0;
1342 if (!arg_a)
1344 in.class = CLASS_ZERO;
1346 else
1348 USItype uarg;
1349 int shift;
1350 in.normal_exp = FRACBITS + NGARDS;
1351 if (in.sign)
1353 /* Special case for minint, since there is no +ve integer
1354 representation for it */
1355 if (arg_a == (- MAX_SI_INT - 1))
1357 return (FLO_type)(- MAX_SI_INT - 1);
1359 uarg = (-arg_a);
1361 else
1362 uarg = arg_a;
1364 in.fraction.ll = uarg;
1365 shift = clzusi (uarg) - (BITS_PER_SI - 1 - FRACBITS - NGARDS);
1366 if (shift > 0)
1368 in.fraction.ll <<= shift;
1369 in.normal_exp -= shift;
1372 return pack_d (&in);
1374 #endif /* L_si_to_sf || L_si_to_df */
1376 #if defined(L_usi_to_sf) || defined(L_usi_to_df) || defined(L_usi_to_tf)
1377 FLO_type
1378 usi_to_float (USItype arg_a)
1380 fp_number_type in;
1382 in.sign = 0;
1383 if (!arg_a)
1385 in.class = CLASS_ZERO;
1387 else
1389 int shift;
1390 in.class = CLASS_NUMBER;
1391 in.normal_exp = FRACBITS + NGARDS;
1392 in.fraction.ll = arg_a;
1394 shift = clzusi (arg_a) - (BITS_PER_SI - 1 - FRACBITS - NGARDS);
1395 if (shift < 0)
1397 fractype guard = in.fraction.ll & (((fractype)1 << -shift) - 1);
1398 in.fraction.ll >>= -shift;
1399 in.fraction.ll |= (guard != 0);
1400 in.normal_exp -= shift;
1402 else if (shift > 0)
1404 in.fraction.ll <<= shift;
1405 in.normal_exp -= shift;
1408 return pack_d (&in);
1410 #endif
1412 #if defined(L_sf_to_si) || defined(L_df_to_si) || defined(L_tf_to_si)
1413 SItype
1414 float_to_si (FLO_type arg_a)
1416 fp_number_type a;
1417 SItype tmp;
1418 FLO_union_type au;
1420 au.value = arg_a;
1421 unpack_d (&au, &a);
1423 if (iszero (&a))
1424 return 0;
1425 if (isnan (&a))
1426 return 0;
1427 /* get reasonable MAX_SI_INT... */
1428 if (isinf (&a))
1429 return a.sign ? (-MAX_SI_INT)-1 : MAX_SI_INT;
1430 /* it is a number, but a small one */
1431 if (a.normal_exp < 0)
1432 return 0;
1433 if (a.normal_exp > BITS_PER_SI - 2)
1434 return a.sign ? (-MAX_SI_INT)-1 : MAX_SI_INT;
1435 tmp = a.fraction.ll >> ((FRACBITS + NGARDS) - a.normal_exp);
1436 return a.sign ? (-tmp) : (tmp);
1438 #endif /* L_sf_to_si || L_df_to_si */
1440 #if defined(L_tf_to_usi)
1441 USItype
1442 float_to_usi (FLO_type arg_a)
1444 fp_number_type a;
1445 FLO_union_type au;
1447 au.value = arg_a;
1448 unpack_d (&au, &a);
1450 if (iszero (&a))
1451 return 0;
1452 if (isnan (&a))
1453 return 0;
1454 /* it is a negative number */
1455 if (a.sign)
1456 return 0;
1457 /* get reasonable MAX_USI_INT... */
1458 if (isinf (&a))
1459 return MAX_USI_INT;
1460 /* it is a number, but a small one */
1461 if (a.normal_exp < 0)
1462 return 0;
1463 if (a.normal_exp > BITS_PER_SI - 1)
1464 return MAX_USI_INT;
1465 else if (a.normal_exp > (FRACBITS + NGARDS))
1466 return a.fraction.ll << (a.normal_exp - (FRACBITS + NGARDS));
1467 else
1468 return a.fraction.ll >> ((FRACBITS + NGARDS) - a.normal_exp);
1470 #endif /* L_tf_to_usi */
1472 #if defined(L_negate_sf) || defined(L_negate_df) || defined(L_negate_tf)
1473 FLO_type
1474 negate (FLO_type arg_a)
1476 fp_number_type a;
1477 FLO_union_type au;
1479 au.value = arg_a;
1480 unpack_d (&au, &a);
1482 flip_sign (&a);
1483 return pack_d (&a);
1485 #endif /* L_negate_sf || L_negate_df */
1487 #ifdef FLOAT
1489 #if defined(L_make_sf)
1490 SFtype
1491 __make_fp(fp_class_type class,
1492 unsigned int sign,
1493 int exp,
1494 USItype frac)
1496 fp_number_type in;
1498 in.class = class;
1499 in.sign = sign;
1500 in.normal_exp = exp;
1501 in.fraction.ll = frac;
1502 return pack_d (&in);
1504 #endif /* L_make_sf */
1506 #ifndef FLOAT_ONLY
1508 /* This enables one to build an fp library that supports float but not double.
1509 Otherwise, we would get an undefined reference to __make_dp.
1510 This is needed for some 8-bit ports that can't handle well values that
1511 are 8-bytes in size, so we just don't support double for them at all. */
1513 #if defined(L_sf_to_df)
1514 DFtype
1515 sf_to_df (SFtype arg_a)
1517 fp_number_type in;
1518 FLO_union_type au;
1520 au.value = arg_a;
1521 unpack_d (&au, &in);
1523 return __make_dp (in.class, in.sign, in.normal_exp,
1524 ((UDItype) in.fraction.ll) << F_D_BITOFF);
1526 #endif /* L_sf_to_df */
1528 #if defined(L_sf_to_tf) && defined(TMODES)
1529 TFtype
1530 sf_to_tf (SFtype arg_a)
1532 fp_number_type in;
1533 FLO_union_type au;
1535 au.value = arg_a;
1536 unpack_d (&au, &in);
1538 return __make_tp (in.class, in.sign, in.normal_exp,
1539 ((UTItype) in.fraction.ll) << F_T_BITOFF);
1541 #endif /* L_sf_to_df */
1543 #endif /* ! FLOAT_ONLY */
1544 #endif /* FLOAT */
1546 #ifndef FLOAT
1548 extern SFtype __make_fp (fp_class_type, unsigned int, int, USItype);
1550 #if defined(L_make_df)
1551 DFtype
1552 __make_dp (fp_class_type class, unsigned int sign, int exp, UDItype frac)
1554 fp_number_type in;
1556 in.class = class;
1557 in.sign = sign;
1558 in.normal_exp = exp;
1559 in.fraction.ll = frac;
1560 return pack_d (&in);
1562 #endif /* L_make_df */
1564 #if defined(L_df_to_sf)
1565 SFtype
1566 df_to_sf (DFtype arg_a)
1568 fp_number_type in;
1569 USItype sffrac;
1570 FLO_union_type au;
1572 au.value = arg_a;
1573 unpack_d (&au, &in);
1575 sffrac = in.fraction.ll >> F_D_BITOFF;
1577 /* We set the lowest guard bit in SFFRAC if we discarded any non
1578 zero bits. */
1579 if ((in.fraction.ll & (((USItype) 1 << F_D_BITOFF) - 1)) != 0)
1580 sffrac |= 1;
1582 return __make_fp (in.class, in.sign, in.normal_exp, sffrac);
1584 #endif /* L_df_to_sf */
1586 #if defined(L_df_to_tf) && defined(TMODES) \
1587 && !defined(FLOAT) && !defined(TFLOAT)
1588 TFtype
1589 df_to_tf (DFtype arg_a)
1591 fp_number_type in;
1592 FLO_union_type au;
1594 au.value = arg_a;
1595 unpack_d (&au, &in);
1597 return __make_tp (in.class, in.sign, in.normal_exp,
1598 ((UTItype) in.fraction.ll) << D_T_BITOFF);
1600 #endif /* L_sf_to_df */
1602 #ifdef TFLOAT
1603 #if defined(L_make_tf)
1604 TFtype
1605 __make_tp(fp_class_type class,
1606 unsigned int sign,
1607 int exp,
1608 UTItype frac)
1610 fp_number_type in;
1612 in.class = class;
1613 in.sign = sign;
1614 in.normal_exp = exp;
1615 in.fraction.ll = frac;
1616 return pack_d (&in);
1618 #endif /* L_make_tf */
1620 #if defined(L_tf_to_df)
1621 DFtype
1622 tf_to_df (TFtype arg_a)
1624 fp_number_type in;
1625 UDItype sffrac;
1626 FLO_union_type au;
1628 au.value = arg_a;
1629 unpack_d (&au, &in);
1631 sffrac = in.fraction.ll >> D_T_BITOFF;
1633 /* We set the lowest guard bit in SFFRAC if we discarded any non
1634 zero bits. */
1635 if ((in.fraction.ll & (((UTItype) 1 << D_T_BITOFF) - 1)) != 0)
1636 sffrac |= 1;
1638 return __make_dp (in.class, in.sign, in.normal_exp, sffrac);
1640 #endif /* L_tf_to_df */
1642 #if defined(L_tf_to_sf)
1643 SFtype
1644 tf_to_sf (TFtype arg_a)
1646 fp_number_type in;
1647 USItype sffrac;
1648 FLO_union_type au;
1650 au.value = arg_a;
1651 unpack_d (&au, &in);
1653 sffrac = in.fraction.ll >> F_T_BITOFF;
1655 /* We set the lowest guard bit in SFFRAC if we discarded any non
1656 zero bits. */
1657 if ((in.fraction.ll & (((UTItype) 1 << F_T_BITOFF) - 1)) != 0)
1658 sffrac |= 1;
1660 return __make_fp (in.class, in.sign, in.normal_exp, sffrac);
1662 #endif /* L_tf_to_sf */
1663 #endif /* TFLOAT */
1665 #endif /* ! FLOAT */
1666 #endif /* !EXTENDED_FLOAT_STUBS */