]> git.ipfire.org Git - thirdparty/gcc.git/blob - libgcc/libgcc2.c
i386: Cleanup ix86_expand_{unary|binary}_operator issues
[thirdparty/gcc.git] / libgcc / libgcc2.c
1 /* More subroutines needed by GCC output code on some machines. */
2 /* Compile this one with gcc. */
3 /* Copyright (C) 1989-2023 Free Software Foundation, Inc.
4
5 This file is part of GCC.
6
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.
11
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.
16
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.
20
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/>. */
25
26 #include "tconfig.h"
27 #include "tsystem.h"
28 #include "coretypes.h"
29 #include "tm.h"
30 #include "libgcc_tm.h"
31
32 #ifdef HAVE_GAS_HIDDEN
33 #define ATTRIBUTE_HIDDEN __attribute__ ((__visibility__ ("hidden")))
34 #else
35 #define ATTRIBUTE_HIDDEN
36 #endif
37
38 /* Work out the largest "word" size that we can deal with on this target. */
39 #if MIN_UNITS_PER_WORD > 4
40 # define LIBGCC2_MAX_UNITS_PER_WORD 8
41 #elif (MIN_UNITS_PER_WORD > 2 \
42 || (MIN_UNITS_PER_WORD > 1 && __SIZEOF_LONG_LONG__ > 4))
43 # define LIBGCC2_MAX_UNITS_PER_WORD 4
44 #else
45 # define LIBGCC2_MAX_UNITS_PER_WORD MIN_UNITS_PER_WORD
46 #endif
47
48 /* Work out what word size we are using for this compilation.
49 The value can be set on the command line. */
50 #ifndef LIBGCC2_UNITS_PER_WORD
51 #define LIBGCC2_UNITS_PER_WORD LIBGCC2_MAX_UNITS_PER_WORD
52 #endif
53
54 #if LIBGCC2_UNITS_PER_WORD <= LIBGCC2_MAX_UNITS_PER_WORD
55
56 #include "libgcc2.h"
57 \f
58 #ifdef DECLARE_LIBRARY_RENAMES
59 DECLARE_LIBRARY_RENAMES
60 #endif
61
62 #if defined (L_negdi2)
63 DWtype
64 __negdi2 (DWtype u)
65 {
66 const DWunion uu = {.ll = u};
67 const DWunion w = { {.low = -uu.s.low,
68 .high = -uu.s.high - ((UWtype) -uu.s.low > 0) } };
69
70 return w.ll;
71 }
72 #endif
73
74 #ifdef L_addvsi3
75 Wtype
76 __addvSI3 (Wtype a, Wtype b)
77 {
78 Wtype w;
79
80 if (__builtin_add_overflow (a, b, &w))
81 abort ();
82
83 return w;
84 }
85 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
86 SItype
87 __addvsi3 (SItype a, SItype b)
88 {
89 SItype w;
90
91 if (__builtin_add_overflow (a, b, &w))
92 abort ();
93
94 return w;
95 }
96 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
97 #endif
98 \f
99 #ifdef L_addvdi3
100 DWtype
101 __addvDI3 (DWtype a, DWtype b)
102 {
103 DWtype w;
104
105 if (__builtin_add_overflow (a, b, &w))
106 abort ();
107
108 return w;
109 }
110 #endif
111 \f
112 #ifdef L_subvsi3
113 Wtype
114 __subvSI3 (Wtype a, Wtype b)
115 {
116 Wtype w;
117
118 if (__builtin_sub_overflow (a, b, &w))
119 abort ();
120
121 return w;
122 }
123 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
124 SItype
125 __subvsi3 (SItype a, SItype b)
126 {
127 SItype w;
128
129 if (__builtin_sub_overflow (a, b, &w))
130 abort ();
131
132 return w;
133 }
134 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
135 #endif
136 \f
137 #ifdef L_subvdi3
138 DWtype
139 __subvDI3 (DWtype a, DWtype b)
140 {
141 DWtype w;
142
143 if (__builtin_sub_overflow (a, b, &w))
144 abort ();
145
146 return w;
147 }
148 #endif
149 \f
150 #ifdef L_mulvsi3
151 Wtype
152 __mulvSI3 (Wtype a, Wtype b)
153 {
154 Wtype w;
155
156 if (__builtin_mul_overflow (a, b, &w))
157 abort ();
158
159 return w;
160 }
161 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
162 SItype
163 __mulvsi3 (SItype a, SItype b)
164 {
165 SItype w;
166
167 if (__builtin_mul_overflow (a, b, &w))
168 abort ();
169
170 return w;
171 }
172 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
173 #endif
174 \f
175 #ifdef L_negvsi2
176 Wtype
177 __negvSI2 (Wtype a)
178 {
179 Wtype w;
180
181 if (__builtin_sub_overflow (0, a, &w))
182 abort ();
183
184 return w;
185 }
186 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
187 SItype
188 __negvsi2 (SItype a)
189 {
190 SItype w;
191
192 if (__builtin_sub_overflow (0, a, &w))
193 abort ();
194
195 return w;
196 }
197 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
198 #endif
199 \f
200 #ifdef L_negvdi2
201 DWtype
202 __negvDI2 (DWtype a)
203 {
204 DWtype w;
205
206 if (__builtin_sub_overflow (0, a, &w))
207 abort ();
208
209 return w;
210 }
211 #endif
212 \f
213 #ifdef L_absvsi2
214 Wtype
215 __absvSI2 (Wtype a)
216 {
217 const Wtype v = 0 - (a < 0);
218 Wtype w;
219
220 if (__builtin_add_overflow (a, v, &w))
221 abort ();
222
223 return v ^ w;
224 }
225 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
226 SItype
227 __absvsi2 (SItype a)
228 {
229 const SItype v = 0 - (a < 0);
230 SItype w;
231
232 if (__builtin_add_overflow (a, v, &w))
233 abort ();
234
235 return v ^ w;
236 }
237 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
238 #endif
239 \f
240 #ifdef L_absvdi2
241 DWtype
242 __absvDI2 (DWtype a)
243 {
244 const DWtype v = 0 - (a < 0);
245 DWtype w;
246
247 if (__builtin_add_overflow (a, v, &w))
248 abort ();
249
250 return v ^ w;
251 }
252 #endif
253 \f
254 #ifdef L_mulvdi3
255 DWtype
256 __mulvDI3 (DWtype u, DWtype v)
257 {
258 /* The unchecked multiplication needs 3 Wtype x Wtype multiplications,
259 but the checked multiplication needs only two. */
260 const DWunion uu = {.ll = u};
261 const DWunion vv = {.ll = v};
262
263 if (__builtin_expect (uu.s.high == uu.s.low >> (W_TYPE_SIZE - 1), 1))
264 {
265 /* u fits in a single Wtype. */
266 if (__builtin_expect (vv.s.high == vv.s.low >> (W_TYPE_SIZE - 1), 1))
267 {
268 /* v fits in a single Wtype as well. */
269 /* A single multiplication. No overflow risk. */
270 return (DWtype) uu.s.low * (DWtype) vv.s.low;
271 }
272 else
273 {
274 /* Two multiplications. */
275 DWunion w0 = {.ll = (UDWtype) (UWtype) uu.s.low
276 * (UDWtype) (UWtype) vv.s.low};
277 DWunion w1 = {.ll = (UDWtype) (UWtype) uu.s.low
278 * (UDWtype) (UWtype) vv.s.high};
279
280 if (vv.s.high < 0)
281 w1.s.high -= uu.s.low;
282 if (uu.s.low < 0)
283 w1.ll -= vv.ll;
284 w1.ll += (UWtype) w0.s.high;
285 if (__builtin_expect (w1.s.high == w1.s.low >> (W_TYPE_SIZE - 1), 1))
286 {
287 w0.s.high = w1.s.low;
288 return w0.ll;
289 }
290 }
291 }
292 else
293 {
294 if (__builtin_expect (vv.s.high == vv.s.low >> (W_TYPE_SIZE - 1), 1))
295 {
296 /* v fits into a single Wtype. */
297 /* Two multiplications. */
298 DWunion w0 = {.ll = (UDWtype) (UWtype) uu.s.low
299 * (UDWtype) (UWtype) vv.s.low};
300 DWunion w1 = {.ll = (UDWtype) (UWtype) uu.s.high
301 * (UDWtype) (UWtype) vv.s.low};
302
303 if (uu.s.high < 0)
304 w1.s.high -= vv.s.low;
305 if (vv.s.low < 0)
306 w1.ll -= uu.ll;
307 w1.ll += (UWtype) w0.s.high;
308 if (__builtin_expect (w1.s.high == w1.s.low >> (W_TYPE_SIZE - 1), 1))
309 {
310 w0.s.high = w1.s.low;
311 return w0.ll;
312 }
313 }
314 else
315 {
316 /* A few sign checks and a single multiplication. */
317 if (uu.s.high >= 0)
318 {
319 if (vv.s.high >= 0)
320 {
321 if (uu.s.high == 0 && vv.s.high == 0)
322 {
323 const DWtype w = (UDWtype) (UWtype) uu.s.low
324 * (UDWtype) (UWtype) vv.s.low;
325 if (__builtin_expect (w >= 0, 1))
326 return w;
327 }
328 }
329 else
330 {
331 if (uu.s.high == 0 && vv.s.high == (Wtype) -1)
332 {
333 DWunion ww = {.ll = (UDWtype) (UWtype) uu.s.low
334 * (UDWtype) (UWtype) vv.s.low};
335
336 ww.s.high -= uu.s.low;
337 if (__builtin_expect (ww.s.high < 0, 1))
338 return ww.ll;
339 }
340 }
341 }
342 else
343 {
344 if (vv.s.high >= 0)
345 {
346 if (uu.s.high == (Wtype) -1 && vv.s.high == 0)
347 {
348 DWunion ww = {.ll = (UDWtype) (UWtype) uu.s.low
349 * (UDWtype) (UWtype) vv.s.low};
350
351 ww.s.high -= vv.s.low;
352 if (__builtin_expect (ww.s.high < 0, 1))
353 return ww.ll;
354 }
355 }
356 else
357 {
358 if ((uu.s.high & vv.s.high) == (Wtype) -1
359 && (uu.s.low | vv.s.low) != 0)
360 {
361 DWunion ww = {.ll = (UDWtype) (UWtype) uu.s.low
362 * (UDWtype) (UWtype) vv.s.low};
363
364 ww.s.high -= uu.s.low;
365 ww.s.high -= vv.s.low;
366 if (__builtin_expect (ww.s.high >= 0, 1))
367 return ww.ll;
368 }
369 }
370 }
371 }
372 }
373
374 /* Overflow. */
375 abort ();
376 }
377 #endif
378 \f
379
380 /* Unless shift functions are defined with full ANSI prototypes,
381 parameter b will be promoted to int if shift_count_type is smaller than an int. */
382 #ifdef L_lshrdi3
383 DWtype
384 __lshrdi3 (DWtype u, shift_count_type b)
385 {
386 if (b == 0)
387 return u;
388
389 const DWunion uu = {.ll = u};
390 const shift_count_type bm = W_TYPE_SIZE - b;
391 DWunion w;
392
393 if (bm <= 0)
394 {
395 w.s.high = 0;
396 w.s.low = (UWtype) uu.s.high >> -bm;
397 }
398 else
399 {
400 const UWtype carries = (UWtype) uu.s.high << bm;
401
402 w.s.high = (UWtype) uu.s.high >> b;
403 w.s.low = ((UWtype) uu.s.low >> b) | carries;
404 }
405
406 return w.ll;
407 }
408 #endif
409
410 #ifdef L_ashldi3
411 DWtype
412 __ashldi3 (DWtype u, shift_count_type b)
413 {
414 if (b == 0)
415 return u;
416
417 const DWunion uu = {.ll = u};
418 const shift_count_type bm = W_TYPE_SIZE - b;
419 DWunion w;
420
421 if (bm <= 0)
422 {
423 w.s.low = 0;
424 w.s.high = (UWtype) uu.s.low << -bm;
425 }
426 else
427 {
428 const UWtype carries = (UWtype) uu.s.low >> bm;
429
430 w.s.low = (UWtype) uu.s.low << b;
431 w.s.high = ((UWtype) uu.s.high << b) | carries;
432 }
433
434 return w.ll;
435 }
436 #endif
437
438 #ifdef L_ashrdi3
439 DWtype
440 __ashrdi3 (DWtype u, shift_count_type b)
441 {
442 if (b == 0)
443 return u;
444
445 const DWunion uu = {.ll = u};
446 const shift_count_type bm = W_TYPE_SIZE - b;
447 DWunion w;
448
449 if (bm <= 0)
450 {
451 /* w.s.high = 1..1 or 0..0 */
452 w.s.high = uu.s.high >> (W_TYPE_SIZE - 1);
453 w.s.low = uu.s.high >> -bm;
454 }
455 else
456 {
457 const UWtype carries = (UWtype) uu.s.high << bm;
458
459 w.s.high = uu.s.high >> b;
460 w.s.low = ((UWtype) uu.s.low >> b) | carries;
461 }
462
463 return w.ll;
464 }
465 #endif
466 \f
467 #ifdef L_bswapsi2
468 SItype
469 __bswapsi2 (SItype u)
470 {
471 return ((((u) & 0xff000000u) >> 24)
472 | (((u) & 0x00ff0000u) >> 8)
473 | (((u) & 0x0000ff00u) << 8)
474 | (((u) & 0x000000ffu) << 24));
475 }
476 #endif
477 #ifdef L_bswapdi2
478 DItype
479 __bswapdi2 (DItype u)
480 {
481 return ((((u) & 0xff00000000000000ull) >> 56)
482 | (((u) & 0x00ff000000000000ull) >> 40)
483 | (((u) & 0x0000ff0000000000ull) >> 24)
484 | (((u) & 0x000000ff00000000ull) >> 8)
485 | (((u) & 0x00000000ff000000ull) << 8)
486 | (((u) & 0x0000000000ff0000ull) << 24)
487 | (((u) & 0x000000000000ff00ull) << 40)
488 | (((u) & 0x00000000000000ffull) << 56));
489 }
490 #endif
491 #ifdef L_ffssi2
492 #undef int
493 int
494 __ffsSI2 (UWtype u)
495 {
496 UWtype count;
497
498 if (u == 0)
499 return 0;
500
501 count_trailing_zeros (count, u);
502 return count + 1;
503 }
504 #endif
505 \f
506 #ifdef L_ffsdi2
507 #undef int
508 int
509 __ffsDI2 (DWtype u)
510 {
511 const DWunion uu = {.ll = u};
512 UWtype word, count, add;
513
514 if (uu.s.low != 0)
515 word = uu.s.low, add = 0;
516 else if (uu.s.high != 0)
517 word = uu.s.high, add = W_TYPE_SIZE;
518 else
519 return 0;
520
521 count_trailing_zeros (count, word);
522 return count + add + 1;
523 }
524 #endif
525 \f
526 #ifdef L_muldi3
527 DWtype
528 __muldi3 (DWtype u, DWtype v)
529 {
530 const DWunion uu = {.ll = u};
531 const DWunion vv = {.ll = v};
532 DWunion w = {.ll = __umulsidi3 (uu.s.low, vv.s.low)};
533
534 w.s.high += ((UWtype) uu.s.low * (UWtype) vv.s.high
535 + (UWtype) uu.s.high * (UWtype) vv.s.low);
536
537 return w.ll;
538 }
539 #endif
540 \f
541 #if (defined (L_udivdi3) || defined (L_divdi3) || \
542 defined (L_umoddi3) || defined (L_moddi3))
543 #if defined (sdiv_qrnnd)
544 #define L_udiv_w_sdiv
545 #endif
546 #endif
547
548 #ifdef L_udiv_w_sdiv
549 #if defined (sdiv_qrnnd)
550 #if (defined (L_udivdi3) || defined (L_divdi3) || \
551 defined (L_umoddi3) || defined (L_moddi3))
552 static inline __attribute__ ((__always_inline__))
553 #endif
554 UWtype
555 __udiv_w_sdiv (UWtype *rp, UWtype a1, UWtype a0, UWtype d)
556 {
557 UWtype q, r;
558 UWtype c0, c1, b1;
559
560 if ((Wtype) d >= 0)
561 {
562 if (a1 < d - a1 - (a0 >> (W_TYPE_SIZE - 1)))
563 {
564 /* Dividend, divisor, and quotient are nonnegative. */
565 sdiv_qrnnd (q, r, a1, a0, d);
566 }
567 else
568 {
569 /* Compute c1*2^32 + c0 = a1*2^32 + a0 - 2^31*d. */
570 sub_ddmmss (c1, c0, a1, a0, d >> 1, d << (W_TYPE_SIZE - 1));
571 /* Divide (c1*2^32 + c0) by d. */
572 sdiv_qrnnd (q, r, c1, c0, d);
573 /* Add 2^31 to quotient. */
574 q += (UWtype) 1 << (W_TYPE_SIZE - 1);
575 }
576 }
577 else
578 {
579 b1 = d >> 1; /* d/2, between 2^30 and 2^31 - 1 */
580 c1 = a1 >> 1; /* A/2 */
581 c0 = (a1 << (W_TYPE_SIZE - 1)) + (a0 >> 1);
582
583 if (a1 < b1) /* A < 2^32*b1, so A/2 < 2^31*b1 */
584 {
585 sdiv_qrnnd (q, r, c1, c0, b1); /* (A/2) / (d/2) */
586
587 r = 2*r + (a0 & 1); /* Remainder from A/(2*b1) */
588 if ((d & 1) != 0)
589 {
590 if (r >= q)
591 r = r - q;
592 else if (q - r <= d)
593 {
594 r = r - q + d;
595 q--;
596 }
597 else
598 {
599 r = r - q + 2*d;
600 q -= 2;
601 }
602 }
603 }
604 else if (c1 < b1) /* So 2^31 <= (A/2)/b1 < 2^32 */
605 {
606 c1 = (b1 - 1) - c1;
607 c0 = ~c0; /* logical NOT */
608
609 sdiv_qrnnd (q, r, c1, c0, b1); /* (A/2) / (d/2) */
610
611 q = ~q; /* (A/2)/b1 */
612 r = (b1 - 1) - r;
613
614 r = 2*r + (a0 & 1); /* A/(2*b1) */
615
616 if ((d & 1) != 0)
617 {
618 if (r >= q)
619 r = r - q;
620 else if (q - r <= d)
621 {
622 r = r - q + d;
623 q--;
624 }
625 else
626 {
627 r = r - q + 2*d;
628 q -= 2;
629 }
630 }
631 }
632 else /* Implies c1 = b1 */
633 { /* Hence a1 = d - 1 = 2*b1 - 1 */
634 if (a0 >= -d)
635 {
636 q = -1;
637 r = a0 + d;
638 }
639 else
640 {
641 q = -2;
642 r = a0 + 2*d;
643 }
644 }
645 }
646
647 *rp = r;
648 return q;
649 }
650 #else
651 /* If sdiv_qrnnd doesn't exist, define dummy __udiv_w_sdiv. */
652 UWtype
653 __udiv_w_sdiv (UWtype *rp __attribute__ ((__unused__)),
654 UWtype a1 __attribute__ ((__unused__)),
655 UWtype a0 __attribute__ ((__unused__)),
656 UWtype d __attribute__ ((__unused__)))
657 {
658 return 0;
659 }
660 #endif
661 #endif
662 \f
663 #if (defined (L_udivdi3) || defined (L_divdi3) || \
664 defined (L_umoddi3) || defined (L_moddi3) || \
665 defined (L_divmoddi4))
666 #define L_udivmoddi4
667 #endif
668
669 #ifdef L_clz
670 const UQItype __clz_tab[256] =
671 {
672 0,1,2,2,3,3,3,3,4,4,4,4,4,4,4,4,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,
673 6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,
674 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
675 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
676 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
677 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
678 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
679 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8
680 };
681 #endif
682 \f
683 #ifdef L_clzsi2
684 #undef int
685 int
686 __clzSI2 (UWtype x)
687 {
688 Wtype ret;
689
690 count_leading_zeros (ret, x);
691
692 return ret;
693 }
694 #endif
695 \f
696 #ifdef L_clzdi2
697 #undef int
698 int
699 __clzDI2 (UDWtype x)
700 {
701 const DWunion uu = {.ll = x};
702 UWtype word;
703 Wtype ret, add;
704
705 if (uu.s.high)
706 word = uu.s.high, add = 0;
707 else
708 word = uu.s.low, add = W_TYPE_SIZE;
709
710 count_leading_zeros (ret, word);
711 return ret + add;
712 }
713 #endif
714 \f
715 #ifdef L_ctzsi2
716 #undef int
717 int
718 __ctzSI2 (UWtype x)
719 {
720 Wtype ret;
721
722 count_trailing_zeros (ret, x);
723
724 return ret;
725 }
726 #endif
727 \f
728 #ifdef L_ctzdi2
729 #undef int
730 int
731 __ctzDI2 (UDWtype x)
732 {
733 const DWunion uu = {.ll = x};
734 UWtype word;
735 Wtype ret, add;
736
737 if (uu.s.low)
738 word = uu.s.low, add = 0;
739 else
740 word = uu.s.high, add = W_TYPE_SIZE;
741
742 count_trailing_zeros (ret, word);
743 return ret + add;
744 }
745 #endif
746 \f
747 #ifdef L_clrsbsi2
748 #undef int
749 int
750 __clrsbSI2 (Wtype x)
751 {
752 Wtype ret;
753
754 if (x < 0)
755 x = ~x;
756 if (x == 0)
757 return W_TYPE_SIZE - 1;
758 count_leading_zeros (ret, x);
759 return ret - 1;
760 }
761 #endif
762 \f
763 #ifdef L_clrsbdi2
764 #undef int
765 int
766 __clrsbDI2 (DWtype x)
767 {
768 const DWunion uu = {.ll = x};
769 UWtype word;
770 Wtype ret, add;
771
772 if (uu.s.high == 0)
773 word = uu.s.low, add = W_TYPE_SIZE;
774 else if (uu.s.high == -1)
775 word = ~uu.s.low, add = W_TYPE_SIZE;
776 else if (uu.s.high >= 0)
777 word = uu.s.high, add = 0;
778 else
779 word = ~uu.s.high, add = 0;
780
781 if (word == 0)
782 ret = W_TYPE_SIZE;
783 else
784 count_leading_zeros (ret, word);
785
786 return ret + add - 1;
787 }
788 #endif
789 \f
790 #ifdef L_popcount_tab
791 const UQItype __popcount_tab[256] =
792 {
793 0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,1,2,2,3,2,3,3,4,2,3,3,4,3,4,4,5,
794 1,2,2,3,2,3,3,4,2,3,3,4,3,4,4,5,2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,
795 1,2,2,3,2,3,3,4,2,3,3,4,3,4,4,5,2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,
796 2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,3,4,4,5,4,5,5,6,4,5,5,6,5,6,6,7,
797 1,2,2,3,2,3,3,4,2,3,3,4,3,4,4,5,2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,
798 2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,3,4,4,5,4,5,5,6,4,5,5,6,5,6,6,7,
799 2,3,3,4,3,4,4,5,3,4,4,5,4,5,5,6,3,4,4,5,4,5,5,6,4,5,5,6,5,6,6,7,
800 3,4,4,5,4,5,5,6,4,5,5,6,5,6,6,7,4,5,5,6,5,6,6,7,5,6,6,7,6,7,7,8
801 };
802 #endif
803 \f
804 #if defined(L_popcountsi2) || defined(L_popcountdi2)
805 #define POPCOUNTCST2(x) (((UWtype) x << __CHAR_BIT__) | x)
806 #define POPCOUNTCST4(x) (((UWtype) x << (2 * __CHAR_BIT__)) | x)
807 #define POPCOUNTCST8(x) (((UWtype) x << (4 * __CHAR_BIT__)) | x)
808 #if W_TYPE_SIZE == __CHAR_BIT__
809 #define POPCOUNTCST(x) x
810 #elif W_TYPE_SIZE == 2 * __CHAR_BIT__
811 #define POPCOUNTCST(x) POPCOUNTCST2 (x)
812 #elif W_TYPE_SIZE == 4 * __CHAR_BIT__
813 #define POPCOUNTCST(x) POPCOUNTCST4 (POPCOUNTCST2 (x))
814 #elif W_TYPE_SIZE == 8 * __CHAR_BIT__
815 #define POPCOUNTCST(x) POPCOUNTCST8 (POPCOUNTCST4 (POPCOUNTCST2 (x)))
816 #endif
817 #endif
818 \f
819 #ifdef L_popcountsi2
820 #undef int
821 int
822 __popcountSI2 (UWtype x)
823 {
824 /* Force table lookup on targets like AVR and RL78 which only
825 pretend they have LIBGCC2_UNITS_PER_WORD 4, but actually
826 have 1, and other small word targets. */
827 #if __SIZEOF_INT__ > 2 && defined (POPCOUNTCST) && __CHAR_BIT__ == 8
828 x = x - ((x >> 1) & POPCOUNTCST (0x55));
829 x = (x & POPCOUNTCST (0x33)) + ((x >> 2) & POPCOUNTCST (0x33));
830 x = (x + (x >> 4)) & POPCOUNTCST (0x0F);
831 return (x * POPCOUNTCST (0x01)) >> (W_TYPE_SIZE - __CHAR_BIT__);
832 #else
833 int i, ret = 0;
834
835 for (i = 0; i < W_TYPE_SIZE; i += 8)
836 ret += __popcount_tab[(x >> i) & 0xff];
837
838 return ret;
839 #endif
840 }
841 #endif
842 \f
843 #ifdef L_popcountdi2
844 #undef int
845 int
846 __popcountDI2 (UDWtype x)
847 {
848 /* Force table lookup on targets like AVR and RL78 which only
849 pretend they have LIBGCC2_UNITS_PER_WORD 4, but actually
850 have 1, and other small word targets. */
851 #if __SIZEOF_INT__ > 2 && defined (POPCOUNTCST) && __CHAR_BIT__ == 8
852 const DWunion uu = {.ll = x};
853 UWtype x1 = uu.s.low, x2 = uu.s.high;
854 x1 = x1 - ((x1 >> 1) & POPCOUNTCST (0x55));
855 x2 = x2 - ((x2 >> 1) & POPCOUNTCST (0x55));
856 x1 = (x1 & POPCOUNTCST (0x33)) + ((x1 >> 2) & POPCOUNTCST (0x33));
857 x2 = (x2 & POPCOUNTCST (0x33)) + ((x2 >> 2) & POPCOUNTCST (0x33));
858 x1 = (x1 + (x1 >> 4)) & POPCOUNTCST (0x0F);
859 x2 = (x2 + (x2 >> 4)) & POPCOUNTCST (0x0F);
860 x1 += x2;
861 return (x1 * POPCOUNTCST (0x01)) >> (W_TYPE_SIZE - __CHAR_BIT__);
862 #else
863 int i, ret = 0;
864
865 for (i = 0; i < 2*W_TYPE_SIZE; i += 8)
866 ret += __popcount_tab[(x >> i) & 0xff];
867
868 return ret;
869 #endif
870 }
871 #endif
872 \f
873 #ifdef L_paritysi2
874 #undef int
875 int
876 __paritySI2 (UWtype x)
877 {
878 #if W_TYPE_SIZE > 64
879 # error "fill out the table"
880 #endif
881 #if W_TYPE_SIZE > 32
882 x ^= x >> 32;
883 #endif
884 #if W_TYPE_SIZE > 16
885 x ^= x >> 16;
886 #endif
887 x ^= x >> 8;
888 x ^= x >> 4;
889 x &= 0xf;
890 return (0x6996 >> x) & 1;
891 }
892 #endif
893 \f
894 #ifdef L_paritydi2
895 #undef int
896 int
897 __parityDI2 (UDWtype x)
898 {
899 const DWunion uu = {.ll = x};
900 UWtype nx = uu.s.low ^ uu.s.high;
901
902 #if W_TYPE_SIZE > 64
903 # error "fill out the table"
904 #endif
905 #if W_TYPE_SIZE > 32
906 nx ^= nx >> 32;
907 #endif
908 #if W_TYPE_SIZE > 16
909 nx ^= nx >> 16;
910 #endif
911 nx ^= nx >> 8;
912 nx ^= nx >> 4;
913 nx &= 0xf;
914 return (0x6996 >> nx) & 1;
915 }
916 #endif
917
918 #ifdef L_udivmoddi4
919 #ifdef TARGET_HAS_NO_HW_DIVIDE
920
921 #if (defined (L_udivdi3) || defined (L_divdi3) || \
922 defined (L_umoddi3) || defined (L_moddi3) || \
923 defined (L_divmoddi4))
924 static inline __attribute__ ((__always_inline__))
925 #endif
926 UDWtype
927 __udivmoddi4 (UDWtype n, UDWtype d, UDWtype *rp)
928 {
929 UDWtype q = 0, r = n, y = d;
930 UWtype lz1, lz2, i, k;
931
932 /* Implements align divisor shift dividend method. This algorithm
933 aligns the divisor under the dividend and then perform number of
934 test-subtract iterations which shift the dividend left. Number of
935 iterations is k + 1 where k is the number of bit positions the
936 divisor must be shifted left to align it under the dividend.
937 quotient bits can be saved in the rightmost positions of the dividend
938 as it shifts left on each test-subtract iteration. */
939
940 if (y <= r)
941 {
942 lz1 = __builtin_clzll (d);
943 lz2 = __builtin_clzll (n);
944
945 k = lz1 - lz2;
946 y = (y << k);
947
948 /* Dividend can exceed 2 ^ (width - 1) - 1 but still be less than the
949 aligned divisor. Normal iteration can drops the high order bit
950 of the dividend. Therefore, first test-subtract iteration is a
951 special case, saving its quotient bit in a separate location and
952 not shifting the dividend. */
953 if (r >= y)
954 {
955 r = r - y;
956 q = (1ULL << k);
957 }
958
959 if (k > 0)
960 {
961 y = y >> 1;
962
963 /* k additional iterations where k regular test subtract shift
964 dividend iterations are done. */
965 i = k;
966 do
967 {
968 if (r >= y)
969 r = ((r - y) << 1) + 1;
970 else
971 r = (r << 1);
972 i = i - 1;
973 } while (i != 0);
974
975 /* First quotient bit is combined with the quotient bits resulting
976 from the k regular iterations. */
977 q = q + r;
978 r = r >> k;
979 q = q - (r << k);
980 }
981 }
982
983 if (rp)
984 *rp = r;
985 return q;
986 }
987 #else
988
989 #if (defined (L_udivdi3) || defined (L_divdi3) || \
990 defined (L_umoddi3) || defined (L_moddi3) || \
991 defined (L_divmoddi4))
992 static inline __attribute__ ((__always_inline__))
993 #endif
994 UDWtype
995 __udivmoddi4 (UDWtype n, UDWtype d, UDWtype *rp)
996 {
997 const DWunion nn = {.ll = n};
998 const DWunion dd = {.ll = d};
999 DWunion rr;
1000 UWtype d0, d1, n0, n1, n2;
1001 UWtype q0, q1;
1002 UWtype b, bm;
1003
1004 d0 = dd.s.low;
1005 d1 = dd.s.high;
1006 n0 = nn.s.low;
1007 n1 = nn.s.high;
1008
1009 #if !UDIV_NEEDS_NORMALIZATION
1010 if (d1 == 0)
1011 {
1012 if (d0 > n1)
1013 {
1014 /* 0q = nn / 0D */
1015
1016 udiv_qrnnd (q0, n0, n1, n0, d0);
1017 q1 = 0;
1018
1019 /* Remainder in n0. */
1020 }
1021 else
1022 {
1023 /* qq = NN / 0d */
1024
1025 if (d0 == 0)
1026 d0 = 1 / d0; /* Divide intentionally by zero. */
1027
1028 udiv_qrnnd (q1, n1, 0, n1, d0);
1029 udiv_qrnnd (q0, n0, n1, n0, d0);
1030
1031 /* Remainder in n0. */
1032 }
1033
1034 if (rp != 0)
1035 {
1036 rr.s.low = n0;
1037 rr.s.high = 0;
1038 *rp = rr.ll;
1039 }
1040 }
1041
1042 #else /* UDIV_NEEDS_NORMALIZATION */
1043
1044 if (d1 == 0)
1045 {
1046 if (d0 > n1)
1047 {
1048 /* 0q = nn / 0D */
1049
1050 count_leading_zeros (bm, d0);
1051
1052 if (bm != 0)
1053 {
1054 /* Normalize, i.e. make the most significant bit of the
1055 denominator set. */
1056
1057 d0 = d0 << bm;
1058 n1 = (n1 << bm) | (n0 >> (W_TYPE_SIZE - bm));
1059 n0 = n0 << bm;
1060 }
1061
1062 udiv_qrnnd (q0, n0, n1, n0, d0);
1063 q1 = 0;
1064
1065 /* Remainder in n0 >> bm. */
1066 }
1067 else
1068 {
1069 /* qq = NN / 0d */
1070
1071 if (d0 == 0)
1072 d0 = 1 / d0; /* Divide intentionally by zero. */
1073
1074 count_leading_zeros (bm, d0);
1075
1076 if (bm == 0)
1077 {
1078 /* From (n1 >= d0) /\ (the most significant bit of d0 is set),
1079 conclude (the most significant bit of n1 is set) /\ (the
1080 leading quotient digit q1 = 1).
1081
1082 This special case is necessary, not an optimization.
1083 (Shifts counts of W_TYPE_SIZE are undefined.) */
1084
1085 n1 -= d0;
1086 q1 = 1;
1087 }
1088 else
1089 {
1090 /* Normalize. */
1091
1092 b = W_TYPE_SIZE - bm;
1093
1094 d0 = d0 << bm;
1095 n2 = n1 >> b;
1096 n1 = (n1 << bm) | (n0 >> b);
1097 n0 = n0 << bm;
1098
1099 udiv_qrnnd (q1, n1, n2, n1, d0);
1100 }
1101
1102 /* n1 != d0... */
1103
1104 udiv_qrnnd (q0, n0, n1, n0, d0);
1105
1106 /* Remainder in n0 >> bm. */
1107 }
1108
1109 if (rp != 0)
1110 {
1111 rr.s.low = n0 >> bm;
1112 rr.s.high = 0;
1113 *rp = rr.ll;
1114 }
1115 }
1116 #endif /* UDIV_NEEDS_NORMALIZATION */
1117
1118 else
1119 {
1120 if (d1 > n1)
1121 {
1122 /* 00 = nn / DD */
1123
1124 q0 = 0;
1125 q1 = 0;
1126
1127 /* Remainder in n1n0. */
1128 if (rp != 0)
1129 {
1130 rr.s.low = n0;
1131 rr.s.high = n1;
1132 *rp = rr.ll;
1133 }
1134 }
1135 else
1136 {
1137 /* 0q = NN / dd */
1138
1139 count_leading_zeros (bm, d1);
1140 if (bm == 0)
1141 {
1142 /* From (n1 >= d1) /\ (the most significant bit of d1 is set),
1143 conclude (the most significant bit of n1 is set) /\ (the
1144 quotient digit q0 = 0 or 1).
1145
1146 This special case is necessary, not an optimization. */
1147
1148 /* The condition on the next line takes advantage of that
1149 n1 >= d1 (true due to program flow). */
1150 if (n1 > d1 || n0 >= d0)
1151 {
1152 q0 = 1;
1153 sub_ddmmss (n1, n0, n1, n0, d1, d0);
1154 }
1155 else
1156 q0 = 0;
1157
1158 q1 = 0;
1159
1160 if (rp != 0)
1161 {
1162 rr.s.low = n0;
1163 rr.s.high = n1;
1164 *rp = rr.ll;
1165 }
1166 }
1167 else
1168 {
1169 UWtype m1, m0;
1170 /* Normalize. */
1171
1172 b = W_TYPE_SIZE - bm;
1173
1174 d1 = (d1 << bm) | (d0 >> b);
1175 d0 = d0 << bm;
1176 n2 = n1 >> b;
1177 n1 = (n1 << bm) | (n0 >> b);
1178 n0 = n0 << bm;
1179
1180 udiv_qrnnd (q0, n1, n2, n1, d1);
1181 umul_ppmm (m1, m0, q0, d0);
1182
1183 if (m1 > n1 || (m1 == n1 && m0 > n0))
1184 {
1185 q0--;
1186 sub_ddmmss (m1, m0, m1, m0, d1, d0);
1187 }
1188
1189 q1 = 0;
1190
1191 /* Remainder in (n1n0 - m1m0) >> bm. */
1192 if (rp != 0)
1193 {
1194 sub_ddmmss (n1, n0, n1, n0, m1, m0);
1195 rr.s.low = (n1 << b) | (n0 >> bm);
1196 rr.s.high = n1 >> bm;
1197 *rp = rr.ll;
1198 }
1199 }
1200 }
1201 }
1202
1203 const DWunion ww = {{.low = q0, .high = q1}};
1204 return ww.ll;
1205 }
1206 #endif
1207 #endif
1208
1209 #ifdef L_divdi3
1210 DWtype
1211 __divdi3 (DWtype u, DWtype v)
1212 {
1213 Wtype c = 0;
1214 DWunion uu = {.ll = u};
1215 DWunion vv = {.ll = v};
1216 DWtype w;
1217
1218 if (uu.s.high < 0)
1219 c = ~c,
1220 uu.ll = -uu.ll;
1221 if (vv.s.high < 0)
1222 c = ~c,
1223 vv.ll = -vv.ll;
1224
1225 w = __udivmoddi4 (uu.ll, vv.ll, (UDWtype *) 0);
1226 if (c)
1227 w = -w;
1228
1229 return w;
1230 }
1231 #endif
1232
1233 #ifdef L_moddi3
1234 DWtype
1235 __moddi3 (DWtype u, DWtype v)
1236 {
1237 Wtype c = 0;
1238 DWunion uu = {.ll = u};
1239 DWunion vv = {.ll = v};
1240 DWtype w;
1241
1242 if (uu.s.high < 0)
1243 c = ~c,
1244 uu.ll = -uu.ll;
1245 if (vv.s.high < 0)
1246 vv.ll = -vv.ll;
1247
1248 (void) __udivmoddi4 (uu.ll, vv.ll, (UDWtype*)&w);
1249 if (c)
1250 w = -w;
1251
1252 return w;
1253 }
1254 #endif
1255
1256 #ifdef L_divmoddi4
1257 DWtype
1258 __divmoddi4 (DWtype u, DWtype v, DWtype *rp)
1259 {
1260 Wtype c1 = 0, c2 = 0;
1261 DWunion uu = {.ll = u};
1262 DWunion vv = {.ll = v};
1263 DWtype w;
1264 DWtype r;
1265
1266 if (uu.s.high < 0)
1267 c1 = ~c1, c2 = ~c2,
1268 uu.ll = -uu.ll;
1269 if (vv.s.high < 0)
1270 c1 = ~c1,
1271 vv.ll = -vv.ll;
1272
1273 w = __udivmoddi4 (uu.ll, vv.ll, (UDWtype*)&r);
1274 if (c1)
1275 w = -w;
1276 if (c2)
1277 r = -r;
1278
1279 *rp = r;
1280 return w;
1281 }
1282 #endif
1283
1284 #ifdef L_umoddi3
1285 UDWtype
1286 __umoddi3 (UDWtype u, UDWtype v)
1287 {
1288 UDWtype w;
1289
1290 (void) __udivmoddi4 (u, v, &w);
1291
1292 return w;
1293 }
1294 #endif
1295
1296 #ifdef L_udivdi3
1297 UDWtype
1298 __udivdi3 (UDWtype n, UDWtype d)
1299 {
1300 return __udivmoddi4 (n, d, (UDWtype *) 0);
1301 }
1302 #endif
1303 \f
1304 #if (defined(__BITINT_MAXWIDTH__) \
1305 && (defined(L_mulbitint3) || defined(L_divmodbitint4)))
1306 /* _BitInt support. */
1307
1308 /* If *P is zero or sign extended (the latter only for PREC < 0) from
1309 some narrower _BitInt value, reduce precision. */
1310
1311 static inline __attribute__((__always_inline__)) SItype
1312 bitint_reduce_prec (const UWtype **p, SItype prec)
1313 {
1314 UWtype mslimb;
1315 SItype i;
1316 if (prec < 0)
1317 {
1318 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1319 i = 0;
1320 #else
1321 i = ((USItype) -1 - prec) / W_TYPE_SIZE;
1322 #endif
1323 mslimb = (*p)[i];
1324 if (mslimb & ((UWtype) 1 << (((USItype) -1 - prec) % W_TYPE_SIZE)))
1325 {
1326 SItype n = ((USItype) -prec) % W_TYPE_SIZE;
1327 if (n)
1328 {
1329 mslimb |= ((UWtype) -1 << (((USItype) -1 - prec) % W_TYPE_SIZE));
1330 if (mslimb == (UWtype) -1)
1331 {
1332 prec += n;
1333 if (prec >= -1)
1334 return -2;
1335 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1336 ++p;
1337 #else
1338 --i;
1339 #endif
1340 mslimb = (*p)[i];
1341 n = 0;
1342 }
1343 }
1344 while (mslimb == (UWtype) -1)
1345 {
1346 prec += W_TYPE_SIZE;
1347 if (prec >= -1)
1348 return -2;
1349 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1350 ++p;
1351 #else
1352 --i;
1353 #endif
1354 mslimb = (*p)[i];
1355 }
1356 if (n == 0)
1357 {
1358 if ((Wtype) mslimb >= 0)
1359 {
1360 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1361 --p;
1362 #endif
1363 return prec - 1;
1364 }
1365 }
1366 return prec;
1367 }
1368 else
1369 prec = -prec;
1370 }
1371 else
1372 {
1373 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1374 i = 0;
1375 #else
1376 i = ((USItype) prec - 1) / W_TYPE_SIZE;
1377 #endif
1378 mslimb = (*p)[i];
1379 }
1380 SItype n = ((USItype) prec) % W_TYPE_SIZE;
1381 if (n)
1382 {
1383 mslimb &= ((UWtype) 1 << (((USItype) prec) % W_TYPE_SIZE)) - 1;
1384 if (mslimb == 0)
1385 {
1386 prec -= n;
1387 if (prec == 0)
1388 return 1;
1389 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1390 ++p;
1391 #else
1392 --i;
1393 #endif
1394 mslimb = (*p)[i];
1395 }
1396 }
1397 while (mslimb == 0)
1398 {
1399 prec -= W_TYPE_SIZE;
1400 if (prec == 0)
1401 return 1;
1402 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1403 ++p;
1404 #else
1405 --i;
1406 #endif
1407 mslimb = (*p)[i];
1408 }
1409 return prec;
1410 }
1411
1412 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1413 # define BITINT_INC -1
1414 # define BITINT_END(be, le) (be)
1415 #else
1416 # define BITINT_INC 1
1417 # define BITINT_END(be, le) (le)
1418 #endif
1419
1420 #ifdef L_mulbitint3
1421 /* D = S * L. */
1422
1423 static UWtype
1424 bitint_mul_1 (UWtype *d, const UWtype *s, UWtype l, SItype n)
1425 {
1426 UWtype sv, hi, lo, c = 0;
1427 do
1428 {
1429 sv = *s;
1430 s += BITINT_INC;
1431 umul_ppmm (hi, lo, sv, l);
1432 c = __builtin_add_overflow (lo, c, &lo) + hi;
1433 *d = lo;
1434 d += BITINT_INC;
1435 }
1436 while (--n);
1437 return c;
1438 }
1439
1440 /* D += S * L. */
1441
1442 static UWtype
1443 bitint_addmul_1 (UWtype *d, const UWtype *s, UWtype l, SItype n)
1444 {
1445 UWtype sv, hi, lo, c = 0;
1446 do
1447 {
1448 sv = *s;
1449 s += BITINT_INC;
1450 umul_ppmm (hi, lo, sv, l);
1451 hi += __builtin_add_overflow (lo, *d, &lo);
1452 c = __builtin_add_overflow (lo, c, &lo) + hi;
1453 *d = lo;
1454 d += BITINT_INC;
1455 }
1456 while (--n);
1457 return c;
1458 }
1459
1460 /* If XPREC is positive, it is precision in bits
1461 of an unsigned _BitInt operand (which has XPREC/W_TYPE_SIZE
1462 full limbs and if Xprec%W_TYPE_SIZE one partial limb.
1463 If Xprec is negative, -XPREC is precision in bits
1464 of a signed _BitInt operand. RETPREC should be always
1465 positive. */
1466
1467 void
1468 __mulbitint3 (UWtype *ret, SItype retprec,
1469 const UWtype *u, SItype uprec,
1470 const UWtype *v, SItype vprec)
1471 {
1472 uprec = bitint_reduce_prec (&u, uprec);
1473 vprec = bitint_reduce_prec (&v, vprec);
1474 USItype auprec = uprec < 0 ? -uprec : uprec;
1475 USItype avprec = vprec < 0 ? -vprec : vprec;
1476
1477 /* Prefer non-negative U.
1478 Otherwise make sure V doesn't have higher precision than U. */
1479 if ((uprec < 0 && vprec >= 0)
1480 || (avprec > auprec && !(uprec >= 0 && vprec < 0)))
1481 {
1482 SItype p;
1483 const UWtype *t;
1484 p = uprec; uprec = vprec; vprec = p;
1485 p = auprec; auprec = avprec; avprec = p;
1486 t = u; u = v; v = t;
1487 }
1488
1489 USItype un = auprec / W_TYPE_SIZE;
1490 USItype un2 = (auprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1491 USItype vn = avprec / W_TYPE_SIZE;
1492 USItype vn2 = (avprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1493 USItype retn = ((USItype) retprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1494 USItype retidx, uidx, vidx;
1495 UWtype vv;
1496 /* Indexes of least significant limb. */
1497 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1498 retidx = retn - 1;
1499 uidx = un2 - 1;
1500 vidx = vn2 - 1;
1501 #else
1502 retidx = 0;
1503 uidx = 0;
1504 vidx = 0;
1505 #endif
1506 if (__builtin_expect (auprec <= W_TYPE_SIZE, 0) && vprec < 0)
1507 {
1508 UWtype uu = u[uidx];
1509 if (__builtin_expect (auprec < W_TYPE_SIZE, 0))
1510 uu &= ((UWtype) 1 << (auprec % W_TYPE_SIZE)) - 1;
1511 if (uu == 0)
1512 {
1513 /* 0 * negative would be otherwise mishandled below, so
1514 handle it specially. */
1515 __builtin_memset (ret, 0, retn * sizeof (UWtype));
1516 return;
1517 }
1518 }
1519 vv = v[vidx];
1520 if (__builtin_expect (avprec < W_TYPE_SIZE, 0))
1521 {
1522 if (vprec > 0)
1523 vv &= ((UWtype) 1 << (avprec % W_TYPE_SIZE)) - 1;
1524 else
1525 vv |= (UWtype) -1 << (avprec % W_TYPE_SIZE);
1526 }
1527
1528 USItype n = un > retn ? retn : un;
1529 USItype n2 = n;
1530 USItype retidx2 = retidx + n * BITINT_INC;
1531 UWtype c = 0, uv = 0;
1532 if (n)
1533 c = bitint_mul_1 (ret + retidx, u + uidx, vv, n);
1534 if (retn > un && un2 != un)
1535 {
1536 UWtype hi, lo;
1537 uv = u[uidx + n * BITINT_INC];
1538 if (uprec > 0)
1539 uv &= ((UWtype) 1 << (auprec % W_TYPE_SIZE)) - 1;
1540 else
1541 uv |= (UWtype) -1 << (auprec % W_TYPE_SIZE);
1542 umul_ppmm (hi, lo, uv, vv);
1543 c = __builtin_add_overflow (lo, c, &lo) + hi;
1544 ret[retidx2] = lo;
1545 retidx2 += BITINT_INC;
1546 ++n2;
1547 }
1548 if (retn > un2)
1549 {
1550 if (uprec < 0)
1551 {
1552 while (n2 < retn)
1553 {
1554 if (n2 >= un2 + vn2)
1555 break;
1556 UWtype hi, lo;
1557 umul_ppmm (hi, lo, (UWtype) -1, vv);
1558 c = __builtin_add_overflow (lo, c, &lo) + hi;
1559 ret[retidx2] = lo;
1560 retidx2 += BITINT_INC;
1561 ++n2;
1562 }
1563 }
1564 else
1565 {
1566 ret[retidx2] = c;
1567 retidx2 += BITINT_INC;
1568 ++n2;
1569 }
1570 /* If RET has more limbs than U after precision reduction,
1571 fill in the remaining limbs. */
1572 while (n2 < retn)
1573 {
1574 if (n2 < un2 + vn2 || (uprec ^ vprec) >= 0)
1575 c = 0;
1576 else
1577 c = (UWtype) -1;
1578 ret[retidx2] = c;
1579 retidx2 += BITINT_INC;
1580 ++n2;
1581 }
1582 }
1583 /* N is now number of possibly non-zero limbs in RET (ignoring
1584 limbs above UN2 + VN2 which if any have been finalized already). */
1585 USItype end = vprec < 0 ? un2 + vn2 : vn2;
1586 if (retn > un2 + vn2) retn = un2 + vn2;
1587 if (end > retn) end = retn;
1588 for (USItype m = 1; m < end; ++m)
1589 {
1590 retidx += BITINT_INC;
1591 vidx += BITINT_INC;
1592 if (m < vn2)
1593 {
1594 vv = v[vidx];
1595 if (__builtin_expect (m == vn, 0))
1596 {
1597 if (vprec > 0)
1598 vv &= ((UWtype) 1 << (avprec % W_TYPE_SIZE)) - 1;
1599 else
1600 vv |= (UWtype) -1 << (avprec % W_TYPE_SIZE);
1601 }
1602 }
1603 else
1604 vv = (UWtype) -1;
1605 if (m + n > retn)
1606 n = retn - m;
1607 c = 0;
1608 if (n)
1609 c = bitint_addmul_1 (ret + retidx, u + uidx, vv, n);
1610 n2 = m + n;
1611 retidx2 = retidx + n * BITINT_INC;
1612 if (n2 < retn && un2 != un)
1613 {
1614 UWtype hi, lo;
1615 umul_ppmm (hi, lo, uv, vv);
1616 hi += __builtin_add_overflow (lo, ret[retidx2], &lo);
1617 c = __builtin_add_overflow (lo, c, &lo) + hi;
1618 ret[retidx2] = lo;
1619 retidx2 += BITINT_INC;
1620 ++n2;
1621 }
1622 if (uprec < 0)
1623 while (n2 < retn)
1624 {
1625 UWtype hi, lo;
1626 umul_ppmm (hi, lo, (UWtype) -1, vv);
1627 hi += __builtin_add_overflow (lo, ret[retidx2], &lo);
1628 c = __builtin_add_overflow (lo, c, &lo) + hi;
1629 ret[retidx2] = lo;
1630 retidx2 += BITINT_INC;
1631 ++n2;
1632 }
1633 else if (n2 < retn)
1634 {
1635 ret[retidx2] = c;
1636 retidx2 += BITINT_INC;
1637 }
1638 }
1639 }
1640 #endif
1641
1642 #ifdef L_divmodbitint4
1643 /* D = -S. */
1644
1645 static void
1646 bitint_negate (UWtype *d, const UWtype *s, SItype n)
1647 {
1648 UWtype c = 1;
1649 do
1650 {
1651 UWtype sv = *s, lo;
1652 s += BITINT_INC;
1653 c = __builtin_add_overflow (~sv, c, &lo);
1654 *d = lo;
1655 d += BITINT_INC;
1656 }
1657 while (--n);
1658 }
1659
1660 /* D -= S * L. */
1661
1662 static UWtype
1663 bitint_submul_1 (UWtype *d, const UWtype *s, UWtype l, SItype n)
1664 {
1665 UWtype sv, hi, lo, c = 0;
1666 do
1667 {
1668 sv = *s;
1669 s += BITINT_INC;
1670 umul_ppmm (hi, lo, sv, l);
1671 hi += __builtin_sub_overflow (*d, lo, &lo);
1672 c = __builtin_sub_overflow (lo, c, &lo) + hi;
1673 *d = lo;
1674 d += BITINT_INC;
1675 }
1676 while (--n);
1677 return c;
1678 }
1679
1680 /* If XPREC is positive, it is precision in bits
1681 of an unsigned _BitInt operand (which has XPREC/W_TYPE_SIZE
1682 full limbs and if Xprec%W_TYPE_SIZE one partial limb.
1683 If Xprec is negative, -XPREC is precision in bits
1684 of a signed _BitInt operand. QPREC and RPREC should be
1685 always non-negative. If either Q or R is NULL (at least
1686 one should be non-NULL), then corresponding QPREC or RPREC
1687 should be 0. */
1688
1689 void
1690 __divmodbitint4 (UWtype *q, SItype qprec,
1691 UWtype *r, SItype rprec,
1692 const UWtype *u, SItype uprec,
1693 const UWtype *v, SItype vprec)
1694 {
1695 uprec = bitint_reduce_prec (&u, uprec);
1696 vprec = bitint_reduce_prec (&v, vprec);
1697 USItype auprec = uprec < 0 ? -uprec : uprec;
1698 USItype avprec = vprec < 0 ? -vprec : vprec;
1699 USItype un = (auprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1700 USItype vn = (avprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1701 USItype qn = ((USItype) qprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1702 USItype rn = ((USItype) rprec + W_TYPE_SIZE - 1) / W_TYPE_SIZE;
1703 USItype up = auprec % W_TYPE_SIZE;
1704 USItype vp = avprec % W_TYPE_SIZE;
1705 if (__builtin_expect (un < vn, 0))
1706 {
1707 /* If abs(v) > abs(u), then q is 0 and r is u. */
1708 if (q)
1709 __builtin_memset (q, 0, qn * sizeof (UWtype));
1710 if (r == NULL)
1711 return;
1712 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1713 r += rn - 1;
1714 u += un - 1;
1715 #endif
1716 if (up)
1717 --un;
1718 if (rn < un)
1719 un = rn;
1720 for (rn -= un; un; --un)
1721 {
1722 *r = *u;
1723 r += BITINT_INC;
1724 u += BITINT_INC;
1725 }
1726 if (!rn)
1727 return;
1728 if (up)
1729 {
1730 if (uprec > 0)
1731 *r = *u & (((UWtype) 1 << up) - 1);
1732 else
1733 *r = *u | ((UWtype) -1 << up);
1734 r += BITINT_INC;
1735 if (!--rn)
1736 return;
1737 }
1738 UWtype c = uprec < 0 ? (UWtype) -1 : (UWtype) 0;
1739 for (; rn; --rn)
1740 {
1741 *r = c;
1742 r += BITINT_INC;
1743 }
1744 return;
1745 }
1746 USItype qn2 = un - vn + 1;
1747 if (qn >= qn2)
1748 qn2 = 0;
1749 USItype sz = un + 1 + vn + qn2;
1750 UWtype *buf = __builtin_alloca (sz * sizeof (UWtype));
1751 USItype uidx, vidx;
1752 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1753 uidx = un - 1;
1754 vidx = vn - 1;
1755 #else
1756 uidx = 0;
1757 vidx = 0;
1758 #endif
1759 if (uprec < 0)
1760 bitint_negate (buf + BITINT_END (uidx + 1, 0), u + uidx, un);
1761 else
1762 __builtin_memcpy (buf + BITINT_END (1, 0), u, un * sizeof (UWtype));
1763 if (up)
1764 buf[BITINT_END (1, un - 1)] &= (((UWtype) 1 << up) - 1);
1765 if (vprec < 0)
1766 bitint_negate (buf + un + 1 + vidx, v + vidx, vn);
1767 else
1768 __builtin_memcpy (buf + un + 1, v, vn * sizeof (UWtype));
1769 if (vp)
1770 buf[un + 1 + BITINT_END (0, vn - 1)] &= (((UWtype) 1 << vp) - 1);
1771 UWtype *u2 = buf;
1772 UWtype *v2 = u2 + un + 1;
1773 UWtype *q2 = v2 + vn;
1774 if (!qn2)
1775 q2 = q + BITINT_END (qn - (un - vn + 1), 0);
1776
1777 /* Knuth's algorithm. See also ../gcc/wide-int.cc (divmod_internal_2). */
1778
1779 #ifndef UDIV_NEEDS_NORMALIZATION
1780 /* Handle single limb divisor first. */
1781 if (vn == 1)
1782 {
1783 UWtype vv = v2[0];
1784 if (vv == 0)
1785 vv = 1 / vv; /* Divide intentionally by zero. */
1786 UWtype k = 0;
1787 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1788 for (SItype i = 0; i <= un - 1; ++i)
1789 #else
1790 for (SItype i = un - 1; i >= 0; --i)
1791 #endif
1792 udiv_qrnnd (q2[i], k, k, u2[BITINT_END (i + 1, i)], vv);
1793 if (r != NULL)
1794 r[BITINT_END (rn - 1, 0)] = k;
1795 }
1796 else
1797 #endif
1798 {
1799 SItype s;
1800 #ifdef UDIV_NEEDS_NORMALIZATION
1801 if (vn == 1 && v2[0] == 0)
1802 s = 0;
1803 else
1804 #endif
1805 if (sizeof (0U) == sizeof (UWtype))
1806 s = __builtin_clz (v2[BITINT_END (0, vn - 1)]);
1807 else if (sizeof (0UL) == sizeof (UWtype))
1808 s = __builtin_clzl (v2[BITINT_END (0, vn - 1)]);
1809 else
1810 s = __builtin_clzll (v2[BITINT_END (0, vn - 1)]);
1811 if (s)
1812 {
1813 /* Normalize by shifting v2 left so that it has msb set. */
1814 const SItype n = sizeof (UWtype) * __CHAR_BIT__;
1815 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1816 for (SItype i = 0; i < vn - 1; ++i)
1817 #else
1818 for (SItype i = vn - 1; i > 0; --i)
1819 #endif
1820 v2[i] = (v2[i] << s) | (v2[i - BITINT_INC] >> (n - s));
1821 v2[vidx] = v2[vidx] << s;
1822 /* And shift u2 left by the same amount. */
1823 u2[BITINT_END (0, un)] = u2[BITINT_END (1, un - 1)] >> (n - s);
1824 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1825 for (SItype i = 1; i < un; ++i)
1826 #else
1827 for (SItype i = un - 1; i > 0; --i)
1828 #endif
1829 u2[i] = (u2[i] << s) | (u2[i - BITINT_INC] >> (n - s));
1830 u2[BITINT_END (un, 0)] = u2[BITINT_END (un, 0)] << s;
1831 }
1832 else
1833 u2[BITINT_END (0, un)] = 0;
1834 #ifdef UDIV_NEEDS_NORMALIZATION
1835 /* Handle single limb divisor first. */
1836 if (vn == 1)
1837 {
1838 UWtype vv = v2[0];
1839 if (vv == 0)
1840 vv = 1 / vv; /* Divide intentionally by zero. */
1841 UWtype k = u2[BITINT_END (0, un)];
1842 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1843 for (SItype i = 0; i <= un - 1; ++i)
1844 #else
1845 for (SItype i = un - 1; i >= 0; --i)
1846 #endif
1847 udiv_qrnnd (q2[i], k, k, u2[BITINT_END (i + 1, i)], vv);
1848 if (r != NULL)
1849 r[BITINT_END (rn - 1, 0)] = k >> s;
1850 }
1851 else
1852 #endif
1853 {
1854 UWtype vv1 = v2[BITINT_END (0, vn - 1)];
1855 UWtype vv0 = v2[BITINT_END (1, vn - 2)];
1856 /* Main loop. */
1857 for (SItype j = un - vn; j >= 0; --j)
1858 {
1859 /* Compute estimate in qhat. */
1860 UWtype uv1 = u2[BITINT_END (un - j - vn, j + vn)];
1861 UWtype uv0 = u2[BITINT_END (un - j - vn + 1, j + vn - 1)];
1862 UWtype qhat, rhat, hi, lo, c;
1863 if (uv1 >= vv1)
1864 {
1865 /* udiv_qrnnd doesn't support quotients which don't
1866 fit into UWtype, so subtract from uv1:uv0 vv1
1867 first. */
1868 uv1 -= vv1 + __builtin_sub_overflow (uv0, vv1, &uv0);
1869 udiv_qrnnd (qhat, rhat, uv1, uv0, vv1);
1870 if (!__builtin_add_overflow (rhat, vv1, &rhat))
1871 goto again;
1872 }
1873 else
1874 {
1875 udiv_qrnnd (qhat, rhat, uv1, uv0, vv1);
1876 again:
1877 umul_ppmm (hi, lo, qhat, vv0);
1878 if (hi > rhat
1879 || (hi == rhat
1880 && lo > u2[BITINT_END (un - j - vn + 2,
1881 j + vn - 2)]))
1882 {
1883 --qhat;
1884 if (!__builtin_add_overflow (rhat, vv1, &rhat))
1885 goto again;
1886 }
1887 }
1888
1889 c = bitint_submul_1 (u2 + BITINT_END (un - j, j),
1890 v2 + BITINT_END (vn - 1, 0), qhat, vn);
1891 u2[BITINT_END (un - j - vn, j + vn)] -= c;
1892 /* If we've subtracted too much, decrease qhat and
1893 and add back. */
1894 if ((Wtype) u2[BITINT_END (un - j - vn, j + vn)] < 0)
1895 {
1896 --qhat;
1897 c = 0;
1898 for (USItype i = 0; i < vn; ++i)
1899 {
1900 UWtype s = v2[BITINT_END (vn - 1 - i, i)];
1901 UWtype d = u2[BITINT_END (un - i - j, i + j)];
1902 UWtype c1 = __builtin_add_overflow (d, s, &d);
1903 UWtype c2 = __builtin_add_overflow (d, c, &d);
1904 c = c1 + c2;
1905 u2[BITINT_END (un - i - j, i + j)] = d;
1906 }
1907 u2[BITINT_END (un - j - vn, j + vn)] += c;
1908 }
1909 q2[BITINT_END (un - vn - j, j)] = qhat;
1910 }
1911 if (r != NULL)
1912 {
1913 if (s)
1914 {
1915 const SItype n = sizeof (UWtype) * __CHAR_BIT__;
1916 /* Unnormalize remainder. */
1917 USItype i;
1918 for (i = 0; i < vn && i < rn; ++i)
1919 r[BITINT_END (rn - 1 - i, i)]
1920 = ((u2[BITINT_END (un - i, i)] >> s)
1921 | (u2[BITINT_END (un - i - 1, i + 1)] << (n - s)));
1922 if (i < rn)
1923 r[BITINT_END (rn - vn, vn - 1)]
1924 = u2[BITINT_END (un - vn + 1, vn - 1)] >> s;
1925 }
1926 else if (rn > vn)
1927 __builtin_memcpy (&r[BITINT_END (rn - vn, 0)],
1928 &u2[BITINT_END (un + 1 - vn, 0)],
1929 vn * sizeof (UWtype));
1930 else
1931 __builtin_memcpy (&r[0], &u2[BITINT_END (un + 1 - rn, 0)],
1932 rn * sizeof (UWtype));
1933 }
1934 }
1935 }
1936 if (q != NULL)
1937 {
1938 if ((uprec < 0) ^ (vprec < 0))
1939 {
1940 /* Negative quotient. */
1941 USItype n;
1942 if (un - vn + 1 > qn)
1943 n = qn;
1944 else
1945 n = un - vn + 1;
1946 bitint_negate (q + BITINT_END (qn - 1, 0),
1947 q2 + BITINT_END (un - vn, 0), n);
1948 if (qn > n)
1949 __builtin_memset (q + BITINT_END (0, n), -1,
1950 (qn - n) * sizeof (UWtype));
1951 }
1952 else
1953 {
1954 /* Positive quotient. */
1955 if (qn2)
1956 __builtin_memcpy (q, q2 + BITINT_END (un - vn + 1 - qn, 0),
1957 qn * sizeof (UWtype));
1958 else if (qn > un - vn + 1)
1959 __builtin_memset (q + BITINT_END (0, un - vn + 1), 0,
1960 (qn - (un - vn + 1)) * sizeof (UWtype));
1961 }
1962 }
1963 if (r != NULL)
1964 {
1965 if (uprec < 0)
1966 {
1967 /* Negative remainder. */
1968 bitint_negate (r + BITINT_END (rn - 1, 0),
1969 r + BITINT_END (rn - 1, 0),
1970 rn > vn ? vn : rn);
1971 if (rn > vn)
1972 __builtin_memset (r + BITINT_END (0, vn), -1,
1973 (rn - vn) * sizeof (UWtype));
1974 }
1975 else
1976 {
1977 /* Positive remainder. */
1978 if (rn > vn)
1979 __builtin_memset (r + BITINT_END (0, vn), 0,
1980 (rn - vn) * sizeof (UWtype));
1981 }
1982 }
1983 }
1984 #endif
1985 #endif
1986 \f
1987 #ifdef L_cmpdi2
1988 cmp_return_type
1989 __cmpdi2 (DWtype a, DWtype b)
1990 {
1991 return (a > b) - (a < b) + 1;
1992 }
1993 #endif
1994
1995 #ifdef L_ucmpdi2
1996 cmp_return_type
1997 __ucmpdi2 (UDWtype a, UDWtype b)
1998 {
1999 return (a > b) - (a < b) + 1;
2000 }
2001 #endif
2002 \f
2003 #if defined(L_fixunstfdi) && LIBGCC2_HAS_TF_MODE
2004 UDWtype
2005 __fixunstfDI (TFtype a)
2006 {
2007 if (a < 0)
2008 return 0;
2009
2010 /* Compute high word of result, as a flonum. */
2011 const TFtype b = (a / Wtype_MAXp1_F);
2012 /* Convert that to fixed (but not to DWtype!),
2013 and shift it into the high word. */
2014 UDWtype v = (UWtype) b;
2015 v <<= W_TYPE_SIZE;
2016 /* Remove high part from the TFtype, leaving the low part as flonum. */
2017 a -= (TFtype)v;
2018 /* Convert that to fixed (but not to DWtype!) and add it in.
2019 Sometimes A comes out negative. This is significant, since
2020 A has more bits than a long int does. */
2021 if (a < 0)
2022 v -= (UWtype) (- a);
2023 else
2024 v += (UWtype) a;
2025 return v;
2026 }
2027 #endif
2028
2029 #if defined(L_fixtfdi) && LIBGCC2_HAS_TF_MODE
2030 DWtype
2031 __fixtfdi (TFtype a)
2032 {
2033 if (a < 0)
2034 return - __fixunstfDI (-a);
2035 return __fixunstfDI (a);
2036 }
2037 #endif
2038
2039 #if defined(L_fixunsxfdi) && LIBGCC2_HAS_XF_MODE
2040 UDWtype
2041 __fixunsxfDI (XFtype a)
2042 {
2043 if (a < 0)
2044 return 0;
2045
2046 /* Compute high word of result, as a flonum. */
2047 const XFtype b = (a / Wtype_MAXp1_F);
2048 /* Convert that to fixed (but not to DWtype!),
2049 and shift it into the high word. */
2050 UDWtype v = (UWtype) b;
2051 v <<= W_TYPE_SIZE;
2052 /* Remove high part from the XFtype, leaving the low part as flonum. */
2053 a -= (XFtype)v;
2054 /* Convert that to fixed (but not to DWtype!) and add it in.
2055 Sometimes A comes out negative. This is significant, since
2056 A has more bits than a long int does. */
2057 if (a < 0)
2058 v -= (UWtype) (- a);
2059 else
2060 v += (UWtype) a;
2061 return v;
2062 }
2063 #endif
2064
2065 #if defined(L_fixxfdi) && LIBGCC2_HAS_XF_MODE
2066 DWtype
2067 __fixxfdi (XFtype a)
2068 {
2069 if (a < 0)
2070 return - __fixunsxfDI (-a);
2071 return __fixunsxfDI (a);
2072 }
2073 #endif
2074
2075 #if defined(L_fixunsdfdi) && LIBGCC2_HAS_DF_MODE
2076 UDWtype
2077 __fixunsdfDI (DFtype a)
2078 {
2079 /* Get high part of result. The division here will just moves the radix
2080 point and will not cause any rounding. Then the conversion to integral
2081 type chops result as desired. */
2082 const UWtype hi = a / Wtype_MAXp1_F;
2083
2084 /* Get low part of result. Convert `hi' to floating type and scale it back,
2085 then subtract this from the number being converted. This leaves the low
2086 part. Convert that to integral type. */
2087 const UWtype lo = a - (DFtype) hi * Wtype_MAXp1_F;
2088
2089 /* Assemble result from the two parts. */
2090 return ((UDWtype) hi << W_TYPE_SIZE) | lo;
2091 }
2092 #endif
2093
2094 #if defined(L_fixdfdi) && LIBGCC2_HAS_DF_MODE
2095 DWtype
2096 __fixdfdi (DFtype a)
2097 {
2098 if (a < 0)
2099 return - __fixunsdfDI (-a);
2100 return __fixunsdfDI (a);
2101 }
2102 #endif
2103
2104 #if defined(L_fixunssfdi) && LIBGCC2_HAS_SF_MODE
2105 UDWtype
2106 __fixunssfDI (SFtype a)
2107 {
2108 #if LIBGCC2_HAS_DF_MODE
2109 /* Convert the SFtype to a DFtype, because that is surely not going
2110 to lose any bits. Some day someone else can write a faster version
2111 that avoids converting to DFtype, and verify it really works right. */
2112 const DFtype dfa = a;
2113
2114 /* Get high part of result. The division here will just moves the radix
2115 point and will not cause any rounding. Then the conversion to integral
2116 type chops result as desired. */
2117 const UWtype hi = dfa / Wtype_MAXp1_F;
2118
2119 /* Get low part of result. Convert `hi' to floating type and scale it back,
2120 then subtract this from the number being converted. This leaves the low
2121 part. Convert that to integral type. */
2122 const UWtype lo = dfa - (DFtype) hi * Wtype_MAXp1_F;
2123
2124 /* Assemble result from the two parts. */
2125 return ((UDWtype) hi << W_TYPE_SIZE) | lo;
2126 #elif FLT_MANT_DIG < W_TYPE_SIZE
2127 if (a < 1)
2128 return 0;
2129 if (a < Wtype_MAXp1_F)
2130 return (UWtype)a;
2131 if (a < Wtype_MAXp1_F * Wtype_MAXp1_F)
2132 {
2133 /* Since we know that there are fewer significant bits in the SFmode
2134 quantity than in a word, we know that we can convert out all the
2135 significant bits in one step, and thus avoid losing bits. */
2136
2137 /* ??? This following loop essentially performs frexpf. If we could
2138 use the real libm function, or poke at the actual bits of the fp
2139 format, it would be significantly faster. */
2140
2141 UWtype shift = 0, counter;
2142 SFtype msb;
2143
2144 a /= Wtype_MAXp1_F;
2145 for (counter = W_TYPE_SIZE / 2; counter != 0; counter >>= 1)
2146 {
2147 SFtype counterf = (UWtype)1 << counter;
2148 if (a >= counterf)
2149 {
2150 shift |= counter;
2151 a /= counterf;
2152 }
2153 }
2154
2155 /* Rescale into the range of one word, extract the bits of that
2156 one word, and shift the result into position. */
2157 a *= Wtype_MAXp1_F;
2158 counter = a;
2159 return (DWtype)counter << shift;
2160 }
2161 return -1;
2162 #else
2163 # error
2164 #endif
2165 }
2166 #endif
2167
2168 #if defined(L_fixsfdi) && LIBGCC2_HAS_SF_MODE
2169 DWtype
2170 __fixsfdi (SFtype a)
2171 {
2172 if (a < 0)
2173 return - __fixunssfDI (-a);
2174 return __fixunssfDI (a);
2175 }
2176 #endif
2177
2178 #if defined(L_floatdixf) && LIBGCC2_HAS_XF_MODE
2179 XFtype
2180 __floatdixf (DWtype u)
2181 {
2182 #if W_TYPE_SIZE > __LIBGCC_XF_MANT_DIG__
2183 # error
2184 #endif
2185 XFtype d = (Wtype) (u >> W_TYPE_SIZE);
2186 d *= Wtype_MAXp1_F;
2187 d += (UWtype)u;
2188 return d;
2189 }
2190 #endif
2191
2192 #if defined(L_floatundixf) && LIBGCC2_HAS_XF_MODE
2193 XFtype
2194 __floatundixf (UDWtype u)
2195 {
2196 #if W_TYPE_SIZE > __LIBGCC_XF_MANT_DIG__
2197 # error
2198 #endif
2199 XFtype d = (UWtype) (u >> W_TYPE_SIZE);
2200 d *= Wtype_MAXp1_F;
2201 d += (UWtype)u;
2202 return d;
2203 }
2204 #endif
2205
2206 #if defined(L_floatditf) && LIBGCC2_HAS_TF_MODE
2207 TFtype
2208 __floatditf (DWtype u)
2209 {
2210 #if W_TYPE_SIZE > __LIBGCC_TF_MANT_DIG__
2211 # error
2212 #endif
2213 TFtype d = (Wtype) (u >> W_TYPE_SIZE);
2214 d *= Wtype_MAXp1_F;
2215 d += (UWtype)u;
2216 return d;
2217 }
2218 #endif
2219
2220 #if defined(L_floatunditf) && LIBGCC2_HAS_TF_MODE
2221 TFtype
2222 __floatunditf (UDWtype u)
2223 {
2224 #if W_TYPE_SIZE > __LIBGCC_TF_MANT_DIG__
2225 # error
2226 #endif
2227 TFtype d = (UWtype) (u >> W_TYPE_SIZE);
2228 d *= Wtype_MAXp1_F;
2229 d += (UWtype)u;
2230 return d;
2231 }
2232 #endif
2233
2234 #if (defined(L_floatdisf) && LIBGCC2_HAS_SF_MODE) \
2235 || (defined(L_floatdidf) && LIBGCC2_HAS_DF_MODE)
2236 #define DI_SIZE (W_TYPE_SIZE * 2)
2237 #define F_MODE_OK(SIZE) \
2238 (SIZE < DI_SIZE \
2239 && SIZE > (DI_SIZE - SIZE + FSSIZE) \
2240 && !AVOID_FP_TYPE_CONVERSION(SIZE))
2241 #if defined(L_floatdisf)
2242 #define FUNC __floatdisf
2243 #define FSTYPE SFtype
2244 #define FSSIZE __LIBGCC_SF_MANT_DIG__
2245 #else
2246 #define FUNC __floatdidf
2247 #define FSTYPE DFtype
2248 #define FSSIZE __LIBGCC_DF_MANT_DIG__
2249 #endif
2250
2251 FSTYPE
2252 FUNC (DWtype u)
2253 {
2254 #if FSSIZE >= W_TYPE_SIZE
2255 /* When the word size is small, we never get any rounding error. */
2256 FSTYPE f = (Wtype) (u >> W_TYPE_SIZE);
2257 f *= Wtype_MAXp1_F;
2258 f += (UWtype)u;
2259 return f;
2260 #elif (LIBGCC2_HAS_DF_MODE && F_MODE_OK (__LIBGCC_DF_MANT_DIG__)) \
2261 || (LIBGCC2_HAS_XF_MODE && F_MODE_OK (__LIBGCC_XF_MANT_DIG__)) \
2262 || (LIBGCC2_HAS_TF_MODE && F_MODE_OK (__LIBGCC_TF_MANT_DIG__))
2263
2264 #if (LIBGCC2_HAS_DF_MODE && F_MODE_OK (__LIBGCC_DF_MANT_DIG__))
2265 # define FSIZE __LIBGCC_DF_MANT_DIG__
2266 # define FTYPE DFtype
2267 #elif (LIBGCC2_HAS_XF_MODE && F_MODE_OK (__LIBGCC_XF_MANT_DIG__))
2268 # define FSIZE __LIBGCC_XF_MANT_DIG__
2269 # define FTYPE XFtype
2270 #elif (LIBGCC2_HAS_TF_MODE && F_MODE_OK (__LIBGCC_TF_MANT_DIG__))
2271 # define FSIZE __LIBGCC_TF_MANT_DIG__
2272 # define FTYPE TFtype
2273 #else
2274 # error
2275 #endif
2276
2277 #define REP_BIT ((UDWtype) 1 << (DI_SIZE - FSIZE))
2278
2279 /* Protect against double-rounding error.
2280 Represent any low-order bits, that might be truncated by a bit that
2281 won't be lost. The bit can go in anywhere below the rounding position
2282 of the FSTYPE. A fixed mask and bit position handles all usual
2283 configurations. */
2284 if (! (- ((DWtype) 1 << FSIZE) < u
2285 && u < ((DWtype) 1 << FSIZE)))
2286 {
2287 if ((UDWtype) u & (REP_BIT - 1))
2288 {
2289 u &= ~ (REP_BIT - 1);
2290 u |= REP_BIT;
2291 }
2292 }
2293
2294 /* Do the calculation in a wider type so that we don't lose any of
2295 the precision of the high word while multiplying it. */
2296 FTYPE f = (Wtype) (u >> W_TYPE_SIZE);
2297 f *= Wtype_MAXp1_F;
2298 f += (UWtype)u;
2299 return (FSTYPE) f;
2300 #else
2301 #if FSSIZE >= W_TYPE_SIZE - 2
2302 # error
2303 #endif
2304 /* Finally, the word size is larger than the number of bits in the
2305 required FSTYPE, and we've got no suitable wider type. The only
2306 way to avoid double rounding is to special case the
2307 extraction. */
2308
2309 /* If there are no high bits set, fall back to one conversion. */
2310 if ((Wtype)u == u)
2311 return (FSTYPE)(Wtype)u;
2312
2313 /* Otherwise, find the power of two. */
2314 Wtype hi = u >> W_TYPE_SIZE;
2315 if (hi < 0)
2316 hi = -(UWtype) hi;
2317
2318 UWtype count, shift;
2319 #if !defined (COUNT_LEADING_ZEROS_0) || COUNT_LEADING_ZEROS_0 != W_TYPE_SIZE
2320 if (hi == 0)
2321 count = W_TYPE_SIZE;
2322 else
2323 #endif
2324 count_leading_zeros (count, hi);
2325
2326 /* No leading bits means u == minimum. */
2327 if (count == 0)
2328 return Wtype_MAXp1_F * (FSTYPE) (hi | ((UWtype) u != 0));
2329
2330 shift = 1 + W_TYPE_SIZE - count;
2331
2332 /* Shift down the most significant bits. */
2333 hi = u >> shift;
2334
2335 /* If we lost any nonzero bits, set the lsb to ensure correct rounding. */
2336 if ((UWtype)u << (W_TYPE_SIZE - shift))
2337 hi |= 1;
2338
2339 /* Convert the one word of data, and rescale. */
2340 FSTYPE f = hi, e;
2341 if (shift == W_TYPE_SIZE)
2342 e = Wtype_MAXp1_F;
2343 /* The following two cases could be merged if we knew that the target
2344 supported a native unsigned->float conversion. More often, we only
2345 have a signed conversion, and have to add extra fixup code. */
2346 else if (shift == W_TYPE_SIZE - 1)
2347 e = Wtype_MAXp1_F / 2;
2348 else
2349 e = (Wtype)1 << shift;
2350 return f * e;
2351 #endif
2352 }
2353 #endif
2354
2355 #if (defined(L_floatundisf) && LIBGCC2_HAS_SF_MODE) \
2356 || (defined(L_floatundidf) && LIBGCC2_HAS_DF_MODE)
2357 #define DI_SIZE (W_TYPE_SIZE * 2)
2358 #define F_MODE_OK(SIZE) \
2359 (SIZE < DI_SIZE \
2360 && SIZE > (DI_SIZE - SIZE + FSSIZE) \
2361 && !AVOID_FP_TYPE_CONVERSION(SIZE))
2362 #if defined(L_floatundisf)
2363 #define FUNC __floatundisf
2364 #define FSTYPE SFtype
2365 #define FSSIZE __LIBGCC_SF_MANT_DIG__
2366 #else
2367 #define FUNC __floatundidf
2368 #define FSTYPE DFtype
2369 #define FSSIZE __LIBGCC_DF_MANT_DIG__
2370 #endif
2371
2372 FSTYPE
2373 FUNC (UDWtype u)
2374 {
2375 #if FSSIZE >= W_TYPE_SIZE
2376 /* When the word size is small, we never get any rounding error. */
2377 FSTYPE f = (UWtype) (u >> W_TYPE_SIZE);
2378 f *= Wtype_MAXp1_F;
2379 f += (UWtype)u;
2380 return f;
2381 #elif (LIBGCC2_HAS_DF_MODE && F_MODE_OK (__LIBGCC_DF_MANT_DIG__)) \
2382 || (LIBGCC2_HAS_XF_MODE && F_MODE_OK (__LIBGCC_XF_MANT_DIG__)) \
2383 || (LIBGCC2_HAS_TF_MODE && F_MODE_OK (__LIBGCC_TF_MANT_DIG__))
2384
2385 #if (LIBGCC2_HAS_DF_MODE && F_MODE_OK (__LIBGCC_DF_MANT_DIG__))
2386 # define FSIZE __LIBGCC_DF_MANT_DIG__
2387 # define FTYPE DFtype
2388 #elif (LIBGCC2_HAS_XF_MODE && F_MODE_OK (__LIBGCC_XF_MANT_DIG__))
2389 # define FSIZE __LIBGCC_XF_MANT_DIG__
2390 # define FTYPE XFtype
2391 #elif (LIBGCC2_HAS_TF_MODE && F_MODE_OK (__LIBGCC_TF_MANT_DIG__))
2392 # define FSIZE __LIBGCC_TF_MANT_DIG__
2393 # define FTYPE TFtype
2394 #else
2395 # error
2396 #endif
2397
2398 #define REP_BIT ((UDWtype) 1 << (DI_SIZE - FSIZE))
2399
2400 /* Protect against double-rounding error.
2401 Represent any low-order bits, that might be truncated by a bit that
2402 won't be lost. The bit can go in anywhere below the rounding position
2403 of the FSTYPE. A fixed mask and bit position handles all usual
2404 configurations. */
2405 if (u >= ((UDWtype) 1 << FSIZE))
2406 {
2407 if ((UDWtype) u & (REP_BIT - 1))
2408 {
2409 u &= ~ (REP_BIT - 1);
2410 u |= REP_BIT;
2411 }
2412 }
2413
2414 /* Do the calculation in a wider type so that we don't lose any of
2415 the precision of the high word while multiplying it. */
2416 FTYPE f = (UWtype) (u >> W_TYPE_SIZE);
2417 f *= Wtype_MAXp1_F;
2418 f += (UWtype)u;
2419 return (FSTYPE) f;
2420 #else
2421 #if FSSIZE == W_TYPE_SIZE - 1
2422 # error
2423 #endif
2424 /* Finally, the word size is larger than the number of bits in the
2425 required FSTYPE, and we've got no suitable wider type. The only
2426 way to avoid double rounding is to special case the
2427 extraction. */
2428
2429 /* If there are no high bits set, fall back to one conversion. */
2430 if ((UWtype)u == u)
2431 return (FSTYPE)(UWtype)u;
2432
2433 /* Otherwise, find the power of two. */
2434 UWtype hi = u >> W_TYPE_SIZE;
2435
2436 UWtype count, shift;
2437 count_leading_zeros (count, hi);
2438
2439 shift = W_TYPE_SIZE - count;
2440
2441 /* Shift down the most significant bits. */
2442 hi = u >> shift;
2443
2444 /* If we lost any nonzero bits, set the lsb to ensure correct rounding. */
2445 if ((UWtype)u << (W_TYPE_SIZE - shift))
2446 hi |= 1;
2447
2448 /* Convert the one word of data, and rescale. */
2449 FSTYPE f = hi, e;
2450 if (shift == W_TYPE_SIZE)
2451 e = Wtype_MAXp1_F;
2452 /* The following two cases could be merged if we knew that the target
2453 supported a native unsigned->float conversion. More often, we only
2454 have a signed conversion, and have to add extra fixup code. */
2455 else if (shift == W_TYPE_SIZE - 1)
2456 e = Wtype_MAXp1_F / 2;
2457 else
2458 e = (Wtype)1 << shift;
2459 return f * e;
2460 #endif
2461 }
2462 #endif
2463
2464 #if defined(L_fixunsxfsi) && LIBGCC2_HAS_XF_MODE
2465 UWtype
2466 __fixunsxfSI (XFtype a)
2467 {
2468 if (a >= - (DFtype) Wtype_MIN)
2469 return (Wtype) (a + Wtype_MIN) - Wtype_MIN;
2470 return (Wtype) a;
2471 }
2472 #endif
2473
2474 #if defined(L_fixunsdfsi) && LIBGCC2_HAS_DF_MODE
2475 UWtype
2476 __fixunsdfSI (DFtype a)
2477 {
2478 if (a >= - (DFtype) Wtype_MIN)
2479 return (Wtype) (a + Wtype_MIN) - Wtype_MIN;
2480 return (Wtype) a;
2481 }
2482 #endif
2483
2484 #if defined(L_fixunssfsi) && LIBGCC2_HAS_SF_MODE
2485 UWtype
2486 __fixunssfSI (SFtype a)
2487 {
2488 if (a >= - (SFtype) Wtype_MIN)
2489 return (Wtype) (a + Wtype_MIN) - Wtype_MIN;
2490 return (Wtype) a;
2491 }
2492 #endif
2493 \f
2494 /* Integer power helper used from __builtin_powi for non-constant
2495 exponents. */
2496
2497 #if (defined(L_powisf2) && LIBGCC2_HAS_SF_MODE) \
2498 || (defined(L_powidf2) && LIBGCC2_HAS_DF_MODE) \
2499 || (defined(L_powixf2) && LIBGCC2_HAS_XF_MODE) \
2500 || (defined(L_powitf2) && LIBGCC2_HAS_TF_MODE)
2501 # if defined(L_powisf2)
2502 # define TYPE SFtype
2503 # define NAME __powisf2
2504 # elif defined(L_powidf2)
2505 # define TYPE DFtype
2506 # define NAME __powidf2
2507 # elif defined(L_powixf2)
2508 # define TYPE XFtype
2509 # define NAME __powixf2
2510 # elif defined(L_powitf2)
2511 # define TYPE TFtype
2512 # define NAME __powitf2
2513 # endif
2514
2515 #undef int
2516 #undef unsigned
2517 TYPE
2518 NAME (TYPE x, int m)
2519 {
2520 unsigned int n = m < 0 ? -(unsigned int) m : (unsigned int) m;
2521 TYPE y = n % 2 ? x : 1;
2522 while (n >>= 1)
2523 {
2524 x = x * x;
2525 if (n % 2)
2526 y = y * x;
2527 }
2528 return m < 0 ? 1/y : y;
2529 }
2530
2531 #endif
2532 \f
2533 #if((defined(L_mulhc3) || defined(L_divhc3)) && LIBGCC2_HAS_HF_MODE) \
2534 || ((defined(L_mulsc3) || defined(L_divsc3)) && LIBGCC2_HAS_SF_MODE) \
2535 || ((defined(L_muldc3) || defined(L_divdc3)) && LIBGCC2_HAS_DF_MODE) \
2536 || ((defined(L_mulxc3) || defined(L_divxc3)) && LIBGCC2_HAS_XF_MODE) \
2537 || ((defined(L_multc3) || defined(L_divtc3)) && LIBGCC2_HAS_TF_MODE)
2538
2539 #undef float
2540 #undef double
2541 #undef long
2542
2543 #if defined(L_mulhc3) || defined(L_divhc3)
2544 # define MTYPE HFtype
2545 # define CTYPE HCtype
2546 # define AMTYPE SFtype
2547 # define MODE hc
2548 # define CEXT __LIBGCC_HF_FUNC_EXT__
2549 # define NOTRUNC (!__LIBGCC_HF_EXCESS_PRECISION__)
2550 #elif defined(L_mulsc3) || defined(L_divsc3)
2551 # define MTYPE SFtype
2552 # define CTYPE SCtype
2553 # define AMTYPE DFtype
2554 # define MODE sc
2555 # define CEXT __LIBGCC_SF_FUNC_EXT__
2556 # define NOTRUNC (!__LIBGCC_SF_EXCESS_PRECISION__)
2557 # define RBIG (__LIBGCC_SF_MAX__ / 2)
2558 # define RMIN (__LIBGCC_SF_MIN__)
2559 # define RMIN2 (__LIBGCC_SF_EPSILON__)
2560 # define RMINSCAL (1 / __LIBGCC_SF_EPSILON__)
2561 # define RMAX2 (RBIG * RMIN2)
2562 #elif defined(L_muldc3) || defined(L_divdc3)
2563 # define MTYPE DFtype
2564 # define CTYPE DCtype
2565 # define MODE dc
2566 # define CEXT __LIBGCC_DF_FUNC_EXT__
2567 # define NOTRUNC (!__LIBGCC_DF_EXCESS_PRECISION__)
2568 # define RBIG (__LIBGCC_DF_MAX__ / 2)
2569 # define RMIN (__LIBGCC_DF_MIN__)
2570 # define RMIN2 (__LIBGCC_DF_EPSILON__)
2571 # define RMINSCAL (1 / __LIBGCC_DF_EPSILON__)
2572 # define RMAX2 (RBIG * RMIN2)
2573 #elif defined(L_mulxc3) || defined(L_divxc3)
2574 # define MTYPE XFtype
2575 # define CTYPE XCtype
2576 # define MODE xc
2577 # define CEXT __LIBGCC_XF_FUNC_EXT__
2578 # define NOTRUNC (!__LIBGCC_XF_EXCESS_PRECISION__)
2579 # define RBIG (__LIBGCC_XF_MAX__ / 2)
2580 # define RMIN (__LIBGCC_XF_MIN__)
2581 # define RMIN2 (__LIBGCC_XF_EPSILON__)
2582 # define RMINSCAL (1 / __LIBGCC_XF_EPSILON__)
2583 # define RMAX2 (RBIG * RMIN2)
2584 #elif defined(L_multc3) || defined(L_divtc3)
2585 # define MTYPE TFtype
2586 # define CTYPE TCtype
2587 # define MODE tc
2588 # define CEXT __LIBGCC_TF_FUNC_EXT__
2589 # define NOTRUNC (!__LIBGCC_TF_EXCESS_PRECISION__)
2590 # if __LIBGCC_TF_MANT_DIG__ == 106
2591 # define RBIG (__LIBGCC_DF_MAX__ / 2)
2592 # define RMIN (__LIBGCC_DF_MIN__)
2593 # define RMIN2 (__LIBGCC_DF_EPSILON__)
2594 # define RMINSCAL (1 / __LIBGCC_DF_EPSILON__)
2595 # else
2596 # define RBIG (__LIBGCC_TF_MAX__ / 2)
2597 # define RMIN (__LIBGCC_TF_MIN__)
2598 # define RMIN2 (__LIBGCC_TF_EPSILON__)
2599 # define RMINSCAL (1 / __LIBGCC_TF_EPSILON__)
2600 # endif
2601 # define RMAX2 (RBIG * RMIN2)
2602 #else
2603 # error
2604 #endif
2605
2606 #define CONCAT3(A,B,C) _CONCAT3(A,B,C)
2607 #define _CONCAT3(A,B,C) A##B##C
2608
2609 #define CONCAT2(A,B) _CONCAT2(A,B)
2610 #define _CONCAT2(A,B) A##B
2611
2612 #define isnan(x) __builtin_isnan (x)
2613 #define isfinite(x) __builtin_isfinite (x)
2614 #define isinf(x) __builtin_isinf (x)
2615
2616 #define INFINITY CONCAT2(__builtin_huge_val, CEXT) ()
2617 #define I 1i
2618
2619 /* Helpers to make the following code slightly less gross. */
2620 #define COPYSIGN CONCAT2(__builtin_copysign, CEXT)
2621 #define FABS CONCAT2(__builtin_fabs, CEXT)
2622
2623 /* Verify that MTYPE matches up with CEXT. */
2624 extern void *compile_type_assert[sizeof(INFINITY) == sizeof(MTYPE) ? 1 : -1];
2625
2626 /* Ensure that we've lost any extra precision. */
2627 #if NOTRUNC
2628 # define TRUNC(x)
2629 #else
2630 # define TRUNC(x) __asm__ ("" : "=m"(x) : "m"(x))
2631 #endif
2632
2633 #if defined(L_mulhc3) || defined(L_mulsc3) || defined(L_muldc3) \
2634 || defined(L_mulxc3) || defined(L_multc3)
2635
2636 CTYPE
2637 CONCAT3(__mul,MODE,3) (MTYPE a, MTYPE b, MTYPE c, MTYPE d)
2638 {
2639 MTYPE ac, bd, ad, bc, x, y;
2640 CTYPE res;
2641
2642 ac = a * c;
2643 bd = b * d;
2644 ad = a * d;
2645 bc = b * c;
2646
2647 TRUNC (ac);
2648 TRUNC (bd);
2649 TRUNC (ad);
2650 TRUNC (bc);
2651
2652 x = ac - bd;
2653 y = ad + bc;
2654
2655 if (isnan (x) && isnan (y))
2656 {
2657 /* Recover infinities that computed as NaN + iNaN. */
2658 _Bool recalc = 0;
2659 if (isinf (a) || isinf (b))
2660 {
2661 /* z is infinite. "Box" the infinity and change NaNs in
2662 the other factor to 0. */
2663 a = COPYSIGN (isinf (a) ? 1 : 0, a);
2664 b = COPYSIGN (isinf (b) ? 1 : 0, b);
2665 if (isnan (c)) c = COPYSIGN (0, c);
2666 if (isnan (d)) d = COPYSIGN (0, d);
2667 recalc = 1;
2668 }
2669 if (isinf (c) || isinf (d))
2670 {
2671 /* w is infinite. "Box" the infinity and change NaNs in
2672 the other factor to 0. */
2673 c = COPYSIGN (isinf (c) ? 1 : 0, c);
2674 d = COPYSIGN (isinf (d) ? 1 : 0, d);
2675 if (isnan (a)) a = COPYSIGN (0, a);
2676 if (isnan (b)) b = COPYSIGN (0, b);
2677 recalc = 1;
2678 }
2679 if (!recalc
2680 && (isinf (ac) || isinf (bd)
2681 || isinf (ad) || isinf (bc)))
2682 {
2683 /* Recover infinities from overflow by changing NaNs to 0. */
2684 if (isnan (a)) a = COPYSIGN (0, a);
2685 if (isnan (b)) b = COPYSIGN (0, b);
2686 if (isnan (c)) c = COPYSIGN (0, c);
2687 if (isnan (d)) d = COPYSIGN (0, d);
2688 recalc = 1;
2689 }
2690 if (recalc)
2691 {
2692 x = INFINITY * (a * c - b * d);
2693 y = INFINITY * (a * d + b * c);
2694 }
2695 }
2696
2697 __real__ res = x;
2698 __imag__ res = y;
2699 return res;
2700 }
2701 #endif /* complex multiply */
2702
2703 #if defined(L_divhc3) || defined(L_divsc3) || defined(L_divdc3) \
2704 || defined(L_divxc3) || defined(L_divtc3)
2705
2706 CTYPE
2707 CONCAT3(__div,MODE,3) (MTYPE a, MTYPE b, MTYPE c, MTYPE d)
2708 {
2709 #if defined(L_divhc3) \
2710 || (defined(L_divsc3) && defined(__LIBGCC_HAVE_HWDBL__) )
2711
2712 /* Half precision is handled with float precision.
2713 float is handled with double precision when double precision
2714 hardware is available.
2715 Due to the additional precision, the simple complex divide
2716 method (without Smith's method) is sufficient to get accurate
2717 answers and runs slightly faster than Smith's method. */
2718
2719 AMTYPE aa, bb, cc, dd;
2720 AMTYPE denom;
2721 MTYPE x, y;
2722 CTYPE res;
2723 aa = a;
2724 bb = b;
2725 cc = c;
2726 dd = d;
2727
2728 denom = (cc * cc) + (dd * dd);
2729 x = ((aa * cc) + (bb * dd)) / denom;
2730 y = ((bb * cc) - (aa * dd)) / denom;
2731
2732 #else
2733 MTYPE denom, ratio, x, y;
2734 CTYPE res;
2735
2736 /* double, extended, long double have significant potential
2737 underflow/overflow errors that can be greatly reduced with
2738 a limited number of tests and adjustments. float is handled
2739 the same way when no HW double is available.
2740 */
2741
2742 /* Scale by max(c,d) to reduce chances of denominator overflowing. */
2743 if (FABS (c) < FABS (d))
2744 {
2745 /* Prevent underflow when denominator is near max representable. */
2746 if (FABS (d) >= RBIG)
2747 {
2748 a = a / 2;
2749 b = b / 2;
2750 c = c / 2;
2751 d = d / 2;
2752 }
2753 /* Avoid overflow/underflow issues when c and d are small.
2754 Scaling up helps avoid some underflows.
2755 No new overflow possible since c&d < RMIN2. */
2756 if (FABS (d) < RMIN2)
2757 {
2758 a = a * RMINSCAL;
2759 b = b * RMINSCAL;
2760 c = c * RMINSCAL;
2761 d = d * RMINSCAL;
2762 }
2763 else
2764 {
2765 if (((FABS (a) < RMIN) && (FABS (b) < RMAX2) && (FABS (d) < RMAX2))
2766 || ((FABS (b) < RMIN) && (FABS (a) < RMAX2)
2767 && (FABS (d) < RMAX2)))
2768 {
2769 a = a * RMINSCAL;
2770 b = b * RMINSCAL;
2771 c = c * RMINSCAL;
2772 d = d * RMINSCAL;
2773 }
2774 }
2775 ratio = c / d;
2776 denom = (c * ratio) + d;
2777 /* Choose alternate order of computation if ratio is subnormal. */
2778 if (FABS (ratio) > RMIN)
2779 {
2780 x = ((a * ratio) + b) / denom;
2781 y = ((b * ratio) - a) / denom;
2782 }
2783 else
2784 {
2785 x = ((c * (a / d)) + b) / denom;
2786 y = ((c * (b / d)) - a) / denom;
2787 }
2788 }
2789 else
2790 {
2791 /* Prevent underflow when denominator is near max representable. */
2792 if (FABS (c) >= RBIG)
2793 {
2794 a = a / 2;
2795 b = b / 2;
2796 c = c / 2;
2797 d = d / 2;
2798 }
2799 /* Avoid overflow/underflow issues when both c and d are small.
2800 Scaling up helps avoid some underflows.
2801 No new overflow possible since both c&d are less than RMIN2. */
2802 if (FABS (c) < RMIN2)
2803 {
2804 a = a * RMINSCAL;
2805 b = b * RMINSCAL;
2806 c = c * RMINSCAL;
2807 d = d * RMINSCAL;
2808 }
2809 else
2810 {
2811 if (((FABS (a) < RMIN) && (FABS (b) < RMAX2) && (FABS (c) < RMAX2))
2812 || ((FABS (b) < RMIN) && (FABS (a) < RMAX2)
2813 && (FABS (c) < RMAX2)))
2814 {
2815 a = a * RMINSCAL;
2816 b = b * RMINSCAL;
2817 c = c * RMINSCAL;
2818 d = d * RMINSCAL;
2819 }
2820 }
2821 ratio = d / c;
2822 denom = (d * ratio) + c;
2823 /* Choose alternate order of computation if ratio is subnormal. */
2824 if (FABS (ratio) > RMIN)
2825 {
2826 x = ((b * ratio) + a) / denom;
2827 y = (b - (a * ratio)) / denom;
2828 }
2829 else
2830 {
2831 x = (a + (d * (b / c))) / denom;
2832 y = (b - (d * (a / c))) / denom;
2833 }
2834 }
2835 #endif
2836
2837 /* Recover infinities and zeros that computed as NaN+iNaN; the only
2838 cases are nonzero/zero, infinite/finite, and finite/infinite. */
2839 if (isnan (x) && isnan (y))
2840 {
2841 if (c == 0.0 && d == 0.0 && (!isnan (a) || !isnan (b)))
2842 {
2843 x = COPYSIGN (INFINITY, c) * a;
2844 y = COPYSIGN (INFINITY, c) * b;
2845 }
2846 else if ((isinf (a) || isinf (b)) && isfinite (c) && isfinite (d))
2847 {
2848 a = COPYSIGN (isinf (a) ? 1 : 0, a);
2849 b = COPYSIGN (isinf (b) ? 1 : 0, b);
2850 x = INFINITY * (a * c + b * d);
2851 y = INFINITY * (b * c - a * d);
2852 }
2853 else if ((isinf (c) || isinf (d)) && isfinite (a) && isfinite (b))
2854 {
2855 c = COPYSIGN (isinf (c) ? 1 : 0, c);
2856 d = COPYSIGN (isinf (d) ? 1 : 0, d);
2857 x = 0.0 * (a * c + b * d);
2858 y = 0.0 * (b * c - a * d);
2859 }
2860 }
2861
2862 __real__ res = x;
2863 __imag__ res = y;
2864 return res;
2865 }
2866 #endif /* complex divide */
2867
2868 #endif /* all complex float routines */
2869 \f
2870 /* From here on down, the routines use normal data types. */
2871
2872 #define SItype bogus_type
2873 #define USItype bogus_type
2874 #define DItype bogus_type
2875 #define UDItype bogus_type
2876 #define SFtype bogus_type
2877 #define DFtype bogus_type
2878 #undef Wtype
2879 #undef UWtype
2880 #undef HWtype
2881 #undef UHWtype
2882 #undef DWtype
2883 #undef UDWtype
2884
2885 #undef char
2886 #undef short
2887 #undef int
2888 #undef long
2889 #undef unsigned
2890 #undef float
2891 #undef double
2892 \f
2893 #ifdef L__gcc_bcmp
2894
2895 /* Like bcmp except the sign is meaningful.
2896 Result is negative if S1 is less than S2,
2897 positive if S1 is greater, 0 if S1 and S2 are equal. */
2898
2899 int
2900 __gcc_bcmp (const unsigned char *s1, const unsigned char *s2, size_t size)
2901 {
2902 while (size > 0)
2903 {
2904 const unsigned char c1 = *s1++, c2 = *s2++;
2905 if (c1 != c2)
2906 return c1 - c2;
2907 size--;
2908 }
2909 return 0;
2910 }
2911
2912 #endif
2913 \f
2914 /* __eprintf used to be used by GCC's private version of <assert.h>.
2915 We no longer provide that header, but this routine remains in libgcc.a
2916 for binary backward compatibility. Note that it is not included in
2917 the shared version of libgcc. */
2918 #ifdef L_eprintf
2919 #ifndef inhibit_libc
2920
2921 #undef NULL /* Avoid errors if stdio.h and our stddef.h mismatch. */
2922 #include <stdio.h>
2923
2924 void
2925 __eprintf (const char *string, const char *expression,
2926 unsigned int line, const char *filename)
2927 {
2928 fprintf (stderr, string, expression, line, filename);
2929 fflush (stderr);
2930 abort ();
2931 }
2932
2933 #endif
2934 #endif
2935
2936 \f
2937 #ifdef L_clear_cache
2938 /* Clear part of an instruction cache. */
2939
2940 void
2941 __clear_cache (void *beg __attribute__((__unused__)),
2942 void *end __attribute__((__unused__)))
2943 {
2944 #ifdef CLEAR_INSN_CACHE
2945 /* Cast the void* pointers to char* as some implementations
2946 of the macro assume the pointers can be subtracted from
2947 one another. */
2948 CLEAR_INSN_CACHE ((char *) beg, (char *) end);
2949 #endif /* CLEAR_INSN_CACHE */
2950 }
2951
2952 #endif /* L_clear_cache */
2953 \f
2954 #ifdef L_trampoline
2955
2956 /* Jump to a trampoline, loading the static chain address. */
2957
2958 #if defined(WINNT) && ! defined(__CYGWIN__)
2959 #define WIN32_LEAN_AND_MEAN
2960 #include <windows.h>
2961 int getpagesize (void);
2962 int mprotect (char *,int, int);
2963
2964 int
2965 getpagesize (void)
2966 {
2967 #ifdef _ALPHA_
2968 return 8192;
2969 #else
2970 return 4096;
2971 #endif
2972 }
2973
2974 int
2975 mprotect (char *addr, int len, int prot)
2976 {
2977 DWORD np, op;
2978
2979 if (prot == 7)
2980 np = 0x40;
2981 else if (prot == 5)
2982 np = 0x20;
2983 else if (prot == 4)
2984 np = 0x10;
2985 else if (prot == 3)
2986 np = 0x04;
2987 else if (prot == 1)
2988 np = 0x02;
2989 else if (prot == 0)
2990 np = 0x01;
2991 else
2992 return -1;
2993
2994 if (VirtualProtect (addr, len, np, &op))
2995 return 0;
2996 else
2997 return -1;
2998 }
2999
3000 #endif /* WINNT && ! __CYGWIN__ */
3001
3002 #ifdef TRANSFER_FROM_TRAMPOLINE
3003 TRANSFER_FROM_TRAMPOLINE
3004 #endif
3005 #endif /* L_trampoline */
3006 \f
3007 #ifndef __CYGWIN__
3008 #ifdef L__main
3009
3010 #include "gbl-ctors.h"
3011
3012 /* Some systems use __main in a way incompatible with its use in gcc, in these
3013 cases use the macros NAME__MAIN to give a quoted symbol and SYMBOL__MAIN to
3014 give the same symbol without quotes for an alternative entry point. You
3015 must define both, or neither. */
3016 #ifndef NAME__MAIN
3017 #define NAME__MAIN "__main"
3018 #define SYMBOL__MAIN __main
3019 #endif
3020
3021 #if defined (__LIBGCC_INIT_SECTION_ASM_OP__) \
3022 || defined (__LIBGCC_INIT_ARRAY_SECTION_ASM_OP__)
3023 #undef HAS_INIT_SECTION
3024 #define HAS_INIT_SECTION
3025 #endif
3026
3027 #if !defined (HAS_INIT_SECTION) || !defined (OBJECT_FORMAT_ELF)
3028
3029 /* Some ELF crosses use crtstuff.c to provide __CTOR_LIST__, but use this
3030 code to run constructors. In that case, we need to handle EH here, too.
3031 But MINGW32 is special because it handles CRTSTUFF and EH on its own. */
3032
3033 #ifdef __MINGW32__
3034 #undef __LIBGCC_EH_FRAME_SECTION_NAME__
3035 #endif
3036
3037 #ifdef __LIBGCC_EH_FRAME_SECTION_NAME__
3038 #include "unwind-dw2-fde.h"
3039 extern unsigned char __EH_FRAME_BEGIN__[];
3040 #endif
3041
3042 /* Run all the global destructors on exit from the program. */
3043
3044 void
3045 __do_global_dtors (void)
3046 {
3047 #ifdef DO_GLOBAL_DTORS_BODY
3048 DO_GLOBAL_DTORS_BODY;
3049 #else
3050 static func_ptr *p = __DTOR_LIST__ + 1;
3051 while (*p)
3052 {
3053 p++;
3054 (*(p-1)) ();
3055 }
3056 #endif
3057 #if defined (__LIBGCC_EH_FRAME_SECTION_NAME__) && !defined (HAS_INIT_SECTION)
3058 {
3059 static int completed = 0;
3060 if (! completed)
3061 {
3062 completed = 1;
3063 __deregister_frame_info (__EH_FRAME_BEGIN__);
3064 }
3065 }
3066 #endif
3067 }
3068 #endif
3069
3070 #ifndef HAS_INIT_SECTION
3071 /* Run all the global constructors on entry to the program. */
3072
3073 void
3074 __do_global_ctors (void)
3075 {
3076 #ifdef __LIBGCC_EH_FRAME_SECTION_NAME__
3077 {
3078 static struct object object;
3079 __register_frame_info (__EH_FRAME_BEGIN__, &object);
3080 }
3081 #endif
3082 DO_GLOBAL_CTORS_BODY;
3083 atexit (__do_global_dtors);
3084 }
3085 #endif /* no HAS_INIT_SECTION */
3086
3087 #if !defined (HAS_INIT_SECTION) || defined (INVOKE__main)
3088 /* Subroutine called automatically by `main'.
3089 Compiling a global function named `main'
3090 produces an automatic call to this function at the beginning.
3091
3092 For many systems, this routine calls __do_global_ctors.
3093 For systems which support a .init section we use the .init section
3094 to run __do_global_ctors, so we need not do anything here. */
3095
3096 extern void SYMBOL__MAIN (void);
3097 void
3098 SYMBOL__MAIN (void)
3099 {
3100 /* Support recursive calls to `main': run initializers just once. */
3101 static int initialized;
3102 if (! initialized)
3103 {
3104 initialized = 1;
3105 __do_global_ctors ();
3106 }
3107 }
3108 #endif /* no HAS_INIT_SECTION or INVOKE__main */
3109
3110 #endif /* L__main */
3111 #endif /* __CYGWIN__ */
3112 \f
3113 #ifdef L_ctors
3114
3115 #include "gbl-ctors.h"
3116
3117 /* Provide default definitions for the lists of constructors and
3118 destructors, so that we don't get linker errors. These symbols are
3119 intentionally bss symbols, so that gld and/or collect will provide
3120 the right values. */
3121
3122 /* We declare the lists here with two elements each,
3123 so that they are valid empty lists if no other definition is loaded.
3124
3125 If we are using the old "set" extensions to have the gnu linker
3126 collect ctors and dtors, then we __CTOR_LIST__ and __DTOR_LIST__
3127 must be in the bss/common section.
3128
3129 Long term no port should use those extensions. But many still do. */
3130 #if !defined(__LIBGCC_INIT_SECTION_ASM_OP__)
3131 #if defined (TARGET_ASM_CONSTRUCTOR) || defined (USE_COLLECT2)
3132 func_ptr __CTOR_LIST__[2] = {0, 0};
3133 func_ptr __DTOR_LIST__[2] = {0, 0};
3134 #else
3135 func_ptr __CTOR_LIST__[2];
3136 func_ptr __DTOR_LIST__[2];
3137 #endif
3138 #endif /* no __LIBGCC_INIT_SECTION_ASM_OP__ */
3139 #endif /* L_ctors */
3140 #endif /* LIBGCC2_UNITS_PER_WORD <= MIN_UNITS_PER_WORD */