1 /* More subroutines needed by GCC output code on some machines. */
2 /* Compile this one with gcc. */
3 /* Copyright (C) 1989-2024 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
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
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/>. */
28 #include "coretypes.h"
30 #include "libgcc_tm.h"
32 #ifdef HAVE_GAS_HIDDEN
33 #define ATTRIBUTE_HIDDEN __attribute__ ((__visibility__ ("hidden")))
35 #define ATTRIBUTE_HIDDEN
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
45 # define LIBGCC2_MAX_UNITS_PER_WORD MIN_UNITS_PER_WORD
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
54 #if LIBGCC2_UNITS_PER_WORD <= LIBGCC2_MAX_UNITS_PER_WORD
58 #ifdef DECLARE_LIBRARY_RENAMES
59 DECLARE_LIBRARY_RENAMES
62 #if defined (L_negdi2)
66 const DWunion uu
= {.ll
= u
};
67 const DWunion w
= { {.low
= -uu
.s
.low
,
68 .high
= -uu
.s
.high
- ((UWtype
) -uu
.s
.low
> 0) } };
76 __addvSI3 (Wtype a
, Wtype b
)
80 if (__builtin_add_overflow (a
, b
, &w
))
85 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
87 __addvsi3 (SItype a
, SItype b
)
91 if (__builtin_add_overflow (a
, b
, &w
))
96 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
101 __addvDI3 (DWtype a
, DWtype b
)
105 if (__builtin_add_overflow (a
, b
, &w
))
114 __subvSI3 (Wtype a
, Wtype b
)
118 if (__builtin_sub_overflow (a
, b
, &w
))
123 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
125 __subvsi3 (SItype a
, SItype b
)
129 if (__builtin_sub_overflow (a
, b
, &w
))
134 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
139 __subvDI3 (DWtype a
, DWtype b
)
143 if (__builtin_sub_overflow (a
, b
, &w
))
152 __mulvSI3 (Wtype a
, Wtype b
)
156 if (__builtin_mul_overflow (a
, b
, &w
))
161 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
163 __mulvsi3 (SItype a
, SItype b
)
167 if (__builtin_mul_overflow (a
, b
, &w
))
172 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
181 if (__builtin_sub_overflow (0, a
, &w
))
186 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
192 if (__builtin_sub_overflow (0, a
, &w
))
197 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
206 if (__builtin_sub_overflow (0, a
, &w
))
217 const Wtype v
= 0 - (a
< 0);
220 if (__builtin_add_overflow (a
, v
, &w
))
225 #ifdef COMPAT_SIMODE_TRAPPING_ARITHMETIC
229 const SItype v
= 0 - (a
< 0);
232 if (__builtin_add_overflow (a
, v
, &w
))
237 #endif /* COMPAT_SIMODE_TRAPPING_ARITHMETIC */
244 const DWtype v
= 0 - (a
< 0);
247 if (__builtin_add_overflow (a
, v
, &w
))
256 __mulvDI3 (DWtype u
, DWtype v
)
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
};
263 if (__builtin_expect (uu
.s
.high
== uu
.s
.low
>> (W_TYPE_SIZE
- 1), 1))
265 /* u fits in a single Wtype. */
266 if (__builtin_expect (vv
.s
.high
== vv
.s
.low
>> (W_TYPE_SIZE
- 1), 1))
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
;
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
};
281 w1
.s
.high
-= uu
.s
.low
;
284 w1
.ll
+= (UWtype
) w0
.s
.high
;
285 if (__builtin_expect (w1
.s
.high
== w1
.s
.low
>> (W_TYPE_SIZE
- 1), 1))
287 w0
.s
.high
= w1
.s
.low
;
294 if (__builtin_expect (vv
.s
.high
== vv
.s
.low
>> (W_TYPE_SIZE
- 1), 1))
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
};
304 w1
.s
.high
-= vv
.s
.low
;
307 w1
.ll
+= (UWtype
) w0
.s
.high
;
308 if (__builtin_expect (w1
.s
.high
== w1
.s
.low
>> (W_TYPE_SIZE
- 1), 1))
310 w0
.s
.high
= w1
.s
.low
;
316 /* A few sign checks and a single multiplication. */
321 if (uu
.s
.high
== 0 && vv
.s
.high
== 0)
323 const DWtype w
= (UDWtype
) (UWtype
) uu
.s
.low
324 * (UDWtype
) (UWtype
) vv
.s
.low
;
325 if (__builtin_expect (w
>= 0, 1))
331 if (uu
.s
.high
== 0 && vv
.s
.high
== (Wtype
) -1)
333 DWunion ww
= {.ll
= (UDWtype
) (UWtype
) uu
.s
.low
334 * (UDWtype
) (UWtype
) vv
.s
.low
};
336 ww
.s
.high
-= uu
.s
.low
;
337 if (__builtin_expect (ww
.s
.high
< 0, 1))
346 if (uu
.s
.high
== (Wtype
) -1 && vv
.s
.high
== 0)
348 DWunion ww
= {.ll
= (UDWtype
) (UWtype
) uu
.s
.low
349 * (UDWtype
) (UWtype
) vv
.s
.low
};
351 ww
.s
.high
-= vv
.s
.low
;
352 if (__builtin_expect (ww
.s
.high
< 0, 1))
358 if ((uu
.s
.high
& vv
.s
.high
) == (Wtype
) -1
359 && (uu
.s
.low
| vv
.s
.low
) != 0)
361 DWunion ww
= {.ll
= (UDWtype
) (UWtype
) uu
.s
.low
362 * (UDWtype
) (UWtype
) vv
.s
.low
};
364 ww
.s
.high
-= uu
.s
.low
;
365 ww
.s
.high
-= vv
.s
.low
;
366 if (__builtin_expect (ww
.s
.high
>= 0, 1))
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. */
384 __lshrdi3 (DWtype u
, shift_count_type b
)
389 const DWunion uu
= {.ll
= u
};
390 const shift_count_type bm
= W_TYPE_SIZE
- b
;
396 w
.s
.low
= (UWtype
) uu
.s
.high
>> -bm
;
400 const UWtype carries
= (UWtype
) uu
.s
.high
<< bm
;
402 w
.s
.high
= (UWtype
) uu
.s
.high
>> b
;
403 w
.s
.low
= ((UWtype
) uu
.s
.low
>> b
) | carries
;
412 __ashldi3 (DWtype u
, shift_count_type b
)
417 const DWunion uu
= {.ll
= u
};
418 const shift_count_type bm
= W_TYPE_SIZE
- b
;
424 w
.s
.high
= (UWtype
) uu
.s
.low
<< -bm
;
428 const UWtype carries
= (UWtype
) uu
.s
.low
>> bm
;
430 w
.s
.low
= (UWtype
) uu
.s
.low
<< b
;
431 w
.s
.high
= ((UWtype
) uu
.s
.high
<< b
) | carries
;
440 __ashrdi3 (DWtype u
, shift_count_type b
)
445 const DWunion uu
= {.ll
= u
};
446 const shift_count_type bm
= W_TYPE_SIZE
- b
;
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
;
457 const UWtype carries
= (UWtype
) uu
.s
.high
<< bm
;
459 w
.s
.high
= uu
.s
.high
>> b
;
460 w
.s
.low
= ((UWtype
) uu
.s
.low
>> b
) | carries
;
469 __bswapsi2 (SItype u
)
471 return ((((u
) & 0xff000000u
) >> 24)
472 | (((u
) & 0x00ff0000u
) >> 8)
473 | (((u
) & 0x0000ff00u
) << 8)
474 | (((u
) & 0x000000ffu
) << 24));
479 __bswapdi2 (DItype u
)
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));
501 count_trailing_zeros (count
, u
);
511 const DWunion uu
= {.ll
= u
};
512 UWtype word
, count
, add
;
515 word
= uu
.s
.low
, add
= 0;
516 else if (uu
.s
.high
!= 0)
517 word
= uu
.s
.high
, add
= W_TYPE_SIZE
;
521 count_trailing_zeros (count
, word
);
522 return count
+ add
+ 1;
528 __muldi3 (DWtype u
, DWtype v
)
530 const DWunion uu
= {.ll
= u
};
531 const DWunion vv
= {.ll
= v
};
532 DWunion w
= {.ll
= __umulsidi3 (uu
.s
.low
, vv
.s
.low
)};
534 w
.s
.high
+= ((UWtype
) uu
.s
.low
* (UWtype
) vv
.s
.high
535 + (UWtype
) uu
.s
.high
* (UWtype
) vv
.s
.low
);
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
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__
))
555 __udiv_w_sdiv (UWtype
*rp
, UWtype a1
, UWtype a0
, UWtype d
)
562 if (a1
< d
- a1
- (a0
>> (W_TYPE_SIZE
- 1)))
564 /* Dividend, divisor, and quotient are nonnegative. */
565 sdiv_qrnnd (q
, r
, a1
, a0
, d
);
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);
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);
583 if (a1
< b1
) /* A < 2^32*b1, so A/2 < 2^31*b1 */
585 sdiv_qrnnd (q
, r
, c1
, c0
, b1
); /* (A/2) / (d/2) */
587 r
= 2*r
+ (a0
& 1); /* Remainder from A/(2*b1) */
604 else if (c1
< b1
) /* So 2^31 <= (A/2)/b1 < 2^32 */
607 c0
= ~c0
; /* logical NOT */
609 sdiv_qrnnd (q
, r
, c1
, c0
, b1
); /* (A/2) / (d/2) */
611 q
= ~q
; /* (A/2)/b1 */
614 r
= 2*r
+ (a0
& 1); /* A/(2*b1) */
632 else /* Implies c1 = b1 */
633 { /* Hence a1 = d - 1 = 2*b1 - 1 */
651 /* If sdiv_qrnnd doesn't exist, define dummy __udiv_w_sdiv. */
653 __udiv_w_sdiv (UWtype
*rp
__attribute__ ((__unused__
)),
654 UWtype a1
__attribute__ ((__unused__
)),
655 UWtype a0
__attribute__ ((__unused__
)),
656 UWtype d
__attribute__ ((__unused__
)))
663 #if (defined (L_udivdi3) || defined (L_divdi3) || \
664 defined (L_umoddi3) || defined (L_moddi3) || \
665 defined (L_divmoddi4))
670 const UQItype __clz_tab
[256] =
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
690 count_leading_zeros (ret
, x
);
701 const DWunion uu
= {.ll
= x
};
706 word
= uu
.s
.high
, add
= 0;
708 word
= uu
.s
.low
, add
= W_TYPE_SIZE
;
710 count_leading_zeros (ret
, word
);
722 count_trailing_zeros (ret
, x
);
733 const DWunion uu
= {.ll
= x
};
738 word
= uu
.s
.low
, add
= 0;
740 word
= uu
.s
.high
, add
= W_TYPE_SIZE
;
742 count_trailing_zeros (ret
, word
);
757 return W_TYPE_SIZE
- 1;
758 count_leading_zeros (ret
, x
);
766 __clrsbDI2 (DWtype x
)
768 const DWunion uu
= {.ll
= x
};
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;
779 word
= ~uu
.s
.high
, add
= 0;
784 count_leading_zeros (ret
, word
);
786 return ret
+ add
- 1;
790 #ifdef L_popcount_tab
791 const UQItype __popcount_tab
[256] =
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
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)))
822 __popcountSI2 (UWtype x
)
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__
);
835 for (i
= 0; i
< W_TYPE_SIZE
; i
+= 8)
836 ret
+= __popcount_tab
[(x
>> i
) & 0xff];
846 __popcountDI2 (UDWtype x
)
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);
861 return (x1
* POPCOUNTCST (0x01)) >> (W_TYPE_SIZE
- __CHAR_BIT__
);
865 for (i
= 0; i
< 2*W_TYPE_SIZE
; i
+= 8)
866 ret
+= __popcount_tab
[(x
>> i
) & 0xff];
876 __paritySI2 (UWtype x
)
879 # error "fill out the table"
890 return (0x6996 >> x
) & 1;
897 __parityDI2 (UDWtype x
)
899 const DWunion uu
= {.ll
= x
};
900 UWtype nx
= uu
.s
.low
^ uu
.s
.high
;
903 # error "fill out the table"
914 return (0x6996 >> nx
) & 1;
919 #ifdef TARGET_HAS_NO_HW_DIVIDE
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__
))
927 __udivmoddi4 (UDWtype n
, UDWtype d
, UDWtype
*rp
)
929 UDWtype q
= 0, r
= n
, y
= d
;
930 UWtype lz1
, lz2
, i
, k
;
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. */
942 lz1
= __builtin_clzll (d
);
943 lz2
= __builtin_clzll (n
);
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. */
963 /* k additional iterations where k regular test subtract shift
964 dividend iterations are done. */
969 r
= ((r
- y
) << 1) + 1;
975 /* First quotient bit is combined with the quotient bits resulting
976 from the k regular iterations. */
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__
))
995 __udivmoddi4 (UDWtype n
, UDWtype d
, UDWtype
*rp
)
997 const DWunion nn
= {.ll
= n
};
998 const DWunion dd
= {.ll
= d
};
1000 UWtype d0
, d1
, n0
, n1
, n2
;
1009 #if !UDIV_NEEDS_NORMALIZATION
1016 udiv_qrnnd (q0
, n0
, n1
, n0
, d0
);
1019 /* Remainder in n0. */
1026 d0
= 1 / d0
; /* Divide intentionally by zero. */
1028 udiv_qrnnd (q1
, n1
, 0, n1
, d0
);
1029 udiv_qrnnd (q0
, n0
, n1
, n0
, d0
);
1031 /* Remainder in n0. */
1042 #else /* UDIV_NEEDS_NORMALIZATION */
1050 count_leading_zeros (bm
, d0
);
1054 /* Normalize, i.e. make the most significant bit of the
1058 n1
= (n1
<< bm
) | (n0
>> (W_TYPE_SIZE
- bm
));
1062 udiv_qrnnd (q0
, n0
, n1
, n0
, d0
);
1065 /* Remainder in n0 >> bm. */
1072 d0
= 1 / d0
; /* Divide intentionally by zero. */
1074 count_leading_zeros (bm
, d0
);
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).
1082 This special case is necessary, not an optimization.
1083 (Shifts counts of W_TYPE_SIZE are undefined.) */
1092 b
= W_TYPE_SIZE
- bm
;
1096 n1
= (n1
<< bm
) | (n0
>> b
);
1099 udiv_qrnnd (q1
, n1
, n2
, n1
, d0
);
1104 udiv_qrnnd (q0
, n0
, n1
, n0
, d0
);
1106 /* Remainder in n0 >> bm. */
1111 rr
.s
.low
= n0
>> bm
;
1116 #endif /* UDIV_NEEDS_NORMALIZATION */
1127 /* Remainder in n1n0. */
1139 count_leading_zeros (bm
, d1
);
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).
1146 This special case is necessary, not an optimization. */
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
)
1153 sub_ddmmss (n1
, n0
, n1
, n0
, d1
, d0
);
1172 b
= W_TYPE_SIZE
- bm
;
1174 d1
= (d1
<< bm
) | (d0
>> b
);
1177 n1
= (n1
<< bm
) | (n0
>> b
);
1180 udiv_qrnnd (q0
, n1
, n2
, n1
, d1
);
1181 umul_ppmm (m1
, m0
, q0
, d0
);
1183 if (m1
> n1
|| (m1
== n1
&& m0
> n0
))
1186 sub_ddmmss (m1
, m0
, m1
, m0
, d1
, d0
);
1191 /* Remainder in (n1n0 - m1m0) >> bm. */
1194 sub_ddmmss (n1
, n0
, n1
, n0
, m1
, m0
);
1195 rr
.s
.low
= (n1
<< b
) | (n0
>> bm
);
1196 rr
.s
.high
= n1
>> bm
;
1203 const DWunion ww
= {{.low
= q0
, .high
= q1
}};
1211 __divdi3 (DWtype u
, DWtype v
)
1214 DWunion uu
= {.ll
= u
};
1215 DWunion vv
= {.ll
= v
};
1225 w
= __udivmoddi4 (uu
.ll
, vv
.ll
, (UDWtype
*) 0);
1235 __moddi3 (DWtype u
, DWtype v
)
1238 DWunion uu
= {.ll
= u
};
1239 DWunion vv
= {.ll
= v
};
1248 (void) __udivmoddi4 (uu
.ll
, vv
.ll
, (UDWtype
*)&w
);
1258 __divmoddi4 (DWtype u
, DWtype v
, DWtype
*rp
)
1260 Wtype c1
= 0, c2
= 0;
1261 DWunion uu
= {.ll
= u
};
1262 DWunion vv
= {.ll
= v
};
1273 w
= __udivmoddi4 (uu
.ll
, vv
.ll
, (UDWtype
*)&r
);
1286 __umoddi3 (UDWtype u
, UDWtype v
)
1290 (void) __udivmoddi4 (u
, v
, &w
);
1298 __udivdi3 (UDWtype n
, UDWtype d
)
1300 return __udivmoddi4 (n
, d
, (UDWtype
*) 0);
1304 #if (defined(__BITINT_MAXWIDTH__) \
1305 && (defined(L_mulbitint3) || defined(L_divmodbitint4)))
1306 /* _BitInt support. */
1308 /* If *P is zero or sign extended (the latter only for PREC < 0) from
1309 some narrower _BitInt value, reduce precision. */
1311 static inline __attribute__((__always_inline__
)) SItype
1312 bitint_reduce_prec (const UBILtype
**p
, SItype prec
)
1318 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1321 i
= ((USItype
) -1 - prec
) / W_TYPE_SIZE
;
1324 if (mslimb
& ((UWtype
) 1 << (((USItype
) -1 - prec
) % W_TYPE_SIZE
)))
1326 SItype n
= ((USItype
) -prec
) % W_TYPE_SIZE
;
1329 mslimb
|= ((UWtype
) -1 << (((USItype
) -1 - prec
) % W_TYPE_SIZE
));
1330 if (mslimb
== (UWtype
) -1)
1335 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1344 while (mslimb
== (UWtype
) -1)
1346 prec
+= W_TYPE_SIZE
;
1349 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1358 if ((Wtype
) mslimb
>= 0)
1360 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1373 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1376 i
= ((USItype
) prec
- 1) / W_TYPE_SIZE
;
1380 SItype n
= ((USItype
) prec
) % W_TYPE_SIZE
;
1383 mslimb
&= ((UWtype
) 1 << (((USItype
) prec
) % W_TYPE_SIZE
)) - 1;
1389 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1399 prec
-= W_TYPE_SIZE
;
1402 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1412 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1413 # define BITINT_INC -1
1414 # define BITINT_END(be, le) (be)
1416 # define BITINT_INC 1
1417 # define BITINT_END(be, le) (le)
1424 bitint_mul_1 (UBILtype
*d
, const UBILtype
*s
, UWtype l
, SItype n
)
1426 UWtype sv
, hi
, lo
, c
= 0;
1431 umul_ppmm (hi
, lo
, sv
, l
);
1432 c
= __builtin_add_overflow (lo
, c
, &lo
) + hi
;
1443 bitint_addmul_1 (UBILtype
*d
, const UBILtype
*s
, UWtype l
, SItype n
)
1445 UWtype sv
, hi
, lo
, c
= 0;
1450 umul_ppmm (hi
, lo
, sv
, l
);
1451 hi
+= __builtin_add_overflow (lo
, *d
, &lo
);
1452 c
= __builtin_add_overflow (lo
, c
, &lo
) + hi
;
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
1468 __mulbitint3 (UBILtype
*ret
, SItype retprec
,
1469 const UBILtype
*u
, SItype uprec
,
1470 const UBILtype
*v
, SItype vprec
)
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
;
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)))
1484 p
= uprec
; uprec
= vprec
; vprec
= p
;
1485 p
= auprec
; auprec
= avprec
; avprec
= p
;
1486 t
= u
; u
= v
; v
= t
;
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
;
1496 /* Indexes of least significant limb. */
1497 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1506 if (__builtin_expect (auprec
<= W_TYPE_SIZE
, 0) && vprec
< 0)
1508 UWtype uu
= u
[uidx
];
1509 if (__builtin_expect (auprec
< W_TYPE_SIZE
, 0))
1510 uu
&= ((UWtype
) 1 << (auprec
% W_TYPE_SIZE
)) - 1;
1513 /* 0 * negative would be otherwise mishandled below, so
1514 handle it specially. */
1515 __builtin_memset (ret
, 0, retn
* sizeof (UWtype
));
1520 if (__builtin_expect (avprec
< W_TYPE_SIZE
, 0))
1523 vv
&= ((UWtype
) 1 << (avprec
% W_TYPE_SIZE
)) - 1;
1525 vv
|= (UWtype
) -1 << (avprec
% W_TYPE_SIZE
);
1528 USItype n
= un
> retn
? retn
: un
;
1530 USItype retidx2
= retidx
+ n
* BITINT_INC
;
1531 UWtype c
= 0, uv
= 0;
1533 c
= bitint_mul_1 (ret
+ retidx
, u
+ uidx
, vv
, n
);
1534 if (retn
> un
&& un2
!= un
)
1537 uv
= u
[uidx
+ n
* BITINT_INC
];
1539 uv
&= ((UWtype
) 1 << (auprec
% W_TYPE_SIZE
)) - 1;
1541 uv
|= (UWtype
) -1 << (auprec
% W_TYPE_SIZE
);
1542 umul_ppmm (hi
, lo
, uv
, vv
);
1543 c
= __builtin_add_overflow (lo
, c
, &lo
) + hi
;
1545 retidx2
+= BITINT_INC
;
1554 if (n2
>= un2
+ vn2
)
1557 umul_ppmm (hi
, lo
, (UWtype
) -1, vv
);
1558 c
= __builtin_add_overflow (lo
, c
, &lo
) + hi
;
1560 retidx2
+= BITINT_INC
;
1567 retidx2
+= BITINT_INC
;
1570 /* If RET has more limbs than U after precision reduction,
1571 fill in the remaining limbs. */
1574 if (n2
< un2
+ vn2
|| (uprec
^ vprec
) >= 0)
1579 retidx2
+= BITINT_INC
;
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
)
1590 retidx
+= BITINT_INC
;
1595 if (__builtin_expect (m
== vn
, 0))
1598 vv
&= ((UWtype
) 1 << (avprec
% W_TYPE_SIZE
)) - 1;
1600 vv
|= (UWtype
) -1 << (avprec
% W_TYPE_SIZE
);
1609 c
= bitint_addmul_1 (ret
+ retidx
, u
+ uidx
, vv
, n
);
1611 retidx2
= retidx
+ n
* BITINT_INC
;
1612 if (n2
< retn
&& un2
!= un
)
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
;
1619 retidx2
+= BITINT_INC
;
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
;
1630 retidx2
+= BITINT_INC
;
1636 retidx2
+= BITINT_INC
;
1642 #ifdef L_divmodbitint4
1646 bitint_negate (UBILtype
*d
, const UBILtype
*s
, SItype n
)
1653 c
= __builtin_add_overflow (~sv
, c
, &lo
);
1663 bitint_submul_1 (UBILtype
*d
, const UBILtype
*s
, UWtype l
, SItype n
)
1665 UWtype sv
, hi
, lo
, c
= 0;
1670 umul_ppmm (hi
, lo
, sv
, l
);
1671 hi
+= __builtin_sub_overflow (*d
, lo
, &lo
);
1672 c
= __builtin_sub_overflow (lo
, c
, &lo
) + hi
;
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
1690 __divmodbitint4 (UBILtype
*q
, SItype qprec
,
1691 UBILtype
*r
, SItype rprec
,
1692 const UBILtype
*u
, SItype uprec
,
1693 const UBILtype
*v
, SItype vprec
)
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))
1707 /* If abs(v) > abs(u), then q is 0 and r is u. */
1709 __builtin_memset (q
, 0, qn
* sizeof (UWtype
));
1712 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1720 for (rn
-= un
; un
; --un
)
1731 *r
= *u
& (((UWtype
) 1 << up
) - 1);
1733 *r
= *u
| ((UWtype
) -1 << up
);
1738 UWtype c
= uprec
< 0 ? (UWtype
) -1 : (UWtype
) 0;
1746 USItype qn2
= un
- vn
+ 1;
1749 USItype sz
= un
+ 1 + vn
+ qn2
;
1750 UBILtype
*buf
= __builtin_alloca (sz
* sizeof (UWtype
));
1752 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1760 bitint_negate (buf
+ BITINT_END (uidx
+ 1, 0), u
+ uidx
, un
);
1762 __builtin_memcpy (buf
+ BITINT_END (1, 0), u
, un
* sizeof (UWtype
));
1764 buf
[BITINT_END (1, un
- 1)] &= (((UWtype
) 1 << up
) - 1);
1766 bitint_negate (buf
+ un
+ 1 + vidx
, v
+ vidx
, vn
);
1768 __builtin_memcpy (buf
+ un
+ 1, v
, vn
* sizeof (UWtype
));
1770 buf
[un
+ 1 + BITINT_END (0, vn
- 1)] &= (((UWtype
) 1 << vp
) - 1);
1772 UBILtype
*v2
= u2
+ un
+ 1;
1773 UBILtype
*q2
= v2
+ vn
;
1775 q2
= q
+ BITINT_END (qn
- (un
- vn
+ 1), 0);
1777 /* Knuth's algorithm. See also ../gcc/wide-int.cc (divmod_internal_2). */
1779 #ifndef UDIV_NEEDS_NORMALIZATION
1780 /* Handle single limb divisor first. */
1785 vv
= 1 / vv
; /* Divide intentionally by zero. */
1787 #if __LIBGCC_BITINT_ORDER__ == __ORDER_BIG_ENDIAN__
1788 for (SItype i
= 0; i
<= un
- 1; ++i
)
1790 for (SItype i
= un
- 1; i
>= 0; --i
)
1792 udiv_qrnnd (q2
[i
], k
, k
, u2
[BITINT_END (i
+ 1, i
)], vv
);
1794 r
[BITINT_END (rn
- 1, 0)] = k
;
1800 #ifdef UDIV_NEEDS_NORMALIZATION
1801 if (vn
== 1 && v2
[0] == 0)
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)]);
1810 s
= __builtin_clzll (v2
[BITINT_END (0, vn
- 1)]);
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
)
1818 for (SItype i
= vn
- 1; i
> 0; --i
)
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
)
1827 for (SItype i
= un
- 1; i
> 0; --i
)
1829 u2
[i
] = (u2
[i
] << s
) | (u2
[i
- BITINT_INC
] >> (n
- s
));
1830 u2
[BITINT_END (un
, 0)] = u2
[BITINT_END (un
, 0)] << s
;
1833 u2
[BITINT_END (0, un
)] = 0;
1834 #ifdef UDIV_NEEDS_NORMALIZATION
1835 /* Handle single limb divisor first. */
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
)
1845 for (SItype i
= un
- 1; i
>= 0; --i
)
1847 udiv_qrnnd (q2
[i
], k
, k
, u2
[BITINT_END (i
+ 1, i
)], vv
);
1849 r
[BITINT_END (rn
- 1, 0)] = k
>> s
;
1854 UWtype vv1
= v2
[BITINT_END (0, vn
- 1)];
1855 UWtype vv0
= v2
[BITINT_END (1, vn
- 2)];
1857 for (SItype j
= un
- vn
; j
>= 0; --j
)
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
;
1865 /* udiv_qrnnd doesn't support quotients which don't
1866 fit into UWtype, so subtract from uv1:uv0 vv1
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
))
1875 udiv_qrnnd (qhat
, rhat
, uv1
, uv0
, vv1
);
1877 umul_ppmm (hi
, lo
, qhat
, vv0
);
1880 && lo
> u2
[BITINT_END (un
- j
- vn
+ 2,
1884 if (!__builtin_add_overflow (rhat
, vv1
, &rhat
))
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
1894 if ((Wtype
) u2
[BITINT_END (un
- j
- vn
, j
+ vn
)] < 0)
1898 for (USItype i
= 0; i
< vn
; ++i
)
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
);
1905 u2
[BITINT_END (un
- i
- j
, i
+ j
)] = d
;
1907 u2
[BITINT_END (un
- j
- vn
, j
+ vn
)] += c
;
1909 q2
[BITINT_END (un
- vn
- j
, j
)] = qhat
;
1915 const SItype n
= sizeof (UWtype
) * __CHAR_BIT__
;
1916 /* Unnormalize remainder. */
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
)));
1923 r
[BITINT_END (rn
- vn
, vn
- 1)]
1924 = u2
[BITINT_END (un
- vn
+ 1, vn
- 1)] >> s
;
1927 __builtin_memcpy (&r
[BITINT_END (rn
- vn
, 0)],
1928 &u2
[BITINT_END (un
+ 1 - vn
, 0)],
1929 vn
* sizeof (UWtype
));
1931 __builtin_memcpy (&r
[0], &u2
[BITINT_END (un
+ 1 - rn
, 0)],
1932 rn
* sizeof (UWtype
));
1938 if ((uprec
< 0) ^ (vprec
< 0))
1940 /* Negative quotient. */
1942 if (un
- vn
+ 1 > qn
)
1946 bitint_negate (q
+ BITINT_END (qn
- 1, 0),
1947 q2
+ BITINT_END (un
- vn
, 0), n
);
1949 __builtin_memset (q
+ BITINT_END (0, n
), -1,
1950 (qn
- n
) * sizeof (UWtype
));
1954 /* Positive quotient. */
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
));
1967 /* Negative remainder. */
1968 bitint_negate (r
+ BITINT_END (rn
- 1, 0),
1969 r
+ BITINT_END (rn
- 1, 0),
1972 __builtin_memset (r
+ BITINT_END (0, vn
), -1,
1973 (rn
- vn
) * sizeof (UWtype
));
1977 /* Positive remainder. */
1979 __builtin_memset (r
+ BITINT_END (0, vn
), 0,
1980 (rn
- vn
) * sizeof (UWtype
));
1989 __cmpdi2 (DWtype a
, DWtype b
)
1991 return (a
> b
) - (a
< b
) + 1;
1997 __ucmpdi2 (UDWtype a
, UDWtype b
)
1999 return (a
> b
) - (a
< b
) + 1;
2003 #if defined(L_fixunstfdi) && LIBGCC2_HAS_TF_MODE
2005 __fixunstfDI (TFtype a
)
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
;
2016 /* Remove high part from the TFtype, leaving the low part as flonum. */
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. */
2022 v
-= (UWtype
) (- a
);
2029 #if defined(L_fixtfdi) && LIBGCC2_HAS_TF_MODE
2031 __fixtfdi (TFtype a
)
2034 return - __fixunstfDI (-a
);
2035 return __fixunstfDI (a
);
2039 #if defined(L_fixunsxfdi) && LIBGCC2_HAS_XF_MODE
2041 __fixunsxfDI (XFtype a
)
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
;
2052 /* Remove high part from the XFtype, leaving the low part as flonum. */
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. */
2058 v
-= (UWtype
) (- a
);
2065 #if defined(L_fixxfdi) && LIBGCC2_HAS_XF_MODE
2067 __fixxfdi (XFtype a
)
2070 return - __fixunsxfDI (-a
);
2071 return __fixunsxfDI (a
);
2075 #if defined(L_fixunsdfdi) && LIBGCC2_HAS_DF_MODE
2077 __fixunsdfDI (DFtype a
)
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
;
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
;
2089 /* Assemble result from the two parts. */
2090 return ((UDWtype
) hi
<< W_TYPE_SIZE
) | lo
;
2094 #if defined(L_fixdfdi) && LIBGCC2_HAS_DF_MODE
2096 __fixdfdi (DFtype a
)
2099 return - __fixunsdfDI (-a
);
2100 return __fixunsdfDI (a
);
2104 #if defined(L_fixunssfdi) && LIBGCC2_HAS_SF_MODE
2106 __fixunssfDI (SFtype a
)
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
;
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
;
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
;
2124 /* Assemble result from the two parts. */
2125 return ((UDWtype
) hi
<< W_TYPE_SIZE
) | lo
;
2126 #elif FLT_MANT_DIG < W_TYPE_SIZE
2129 if (a
< Wtype_MAXp1_F
)
2131 if (a
< Wtype_MAXp1_F
* Wtype_MAXp1_F
)
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. */
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. */
2141 UWtype shift
= 0, counter
;
2145 for (counter
= W_TYPE_SIZE
/ 2; counter
!= 0; counter
>>= 1)
2147 SFtype counterf
= (UWtype
)1 << counter
;
2155 /* Rescale into the range of one word, extract the bits of that
2156 one word, and shift the result into position. */
2159 return (DWtype
)counter
<< shift
;
2168 #if defined(L_fixsfdi) && LIBGCC2_HAS_SF_MODE
2170 __fixsfdi (SFtype a
)
2173 return - __fixunssfDI (-a
);
2174 return __fixunssfDI (a
);
2178 #if defined(L_floatdixf) && LIBGCC2_HAS_XF_MODE
2180 __floatdixf (DWtype u
)
2182 #if W_TYPE_SIZE > __LIBGCC_XF_MANT_DIG__
2185 XFtype d
= (Wtype
) (u
>> W_TYPE_SIZE
);
2192 #if defined(L_floatundixf) && LIBGCC2_HAS_XF_MODE
2194 __floatundixf (UDWtype u
)
2196 #if W_TYPE_SIZE > __LIBGCC_XF_MANT_DIG__
2199 XFtype d
= (UWtype
) (u
>> W_TYPE_SIZE
);
2206 #if defined(L_floatditf) && LIBGCC2_HAS_TF_MODE
2208 __floatditf (DWtype u
)
2210 #if W_TYPE_SIZE > __LIBGCC_TF_MANT_DIG__
2213 TFtype d
= (Wtype
) (u
>> W_TYPE_SIZE
);
2220 #if defined(L_floatunditf) && LIBGCC2_HAS_TF_MODE
2222 __floatunditf (UDWtype u
)
2224 #if W_TYPE_SIZE > __LIBGCC_TF_MANT_DIG__
2227 TFtype d
= (UWtype
) (u
>> W_TYPE_SIZE
);
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) \
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__
2246 #define FUNC __floatdidf
2247 #define FSTYPE DFtype
2248 #define FSSIZE __LIBGCC_DF_MANT_DIG__
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
);
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__))
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
2277 #define REP_BIT ((UDWtype) 1 << (DI_SIZE - FSIZE))
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
2284 if (! (- ((DWtype
) 1 << FSIZE
) < u
2285 && u
< ((DWtype
) 1 << FSIZE
)))
2287 if ((UDWtype
) u
& (REP_BIT
- 1))
2289 u
&= ~ (REP_BIT
- 1);
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
);
2301 #if FSSIZE >= W_TYPE_SIZE - 2
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
2309 /* If there are no high bits set, fall back to one conversion. */
2311 return (FSTYPE
)(Wtype
)u
;
2313 /* Otherwise, find the power of two. */
2314 Wtype hi
= u
>> W_TYPE_SIZE
;
2318 UWtype count
, shift
;
2319 #if !defined (COUNT_LEADING_ZEROS_0) || COUNT_LEADING_ZEROS_0 != W_TYPE_SIZE
2321 count
= W_TYPE_SIZE
;
2324 count_leading_zeros (count
, hi
);
2326 /* No leading bits means u == minimum. */
2328 return Wtype_MAXp1_F
* (FSTYPE
) (hi
| ((UWtype
) u
!= 0));
2330 shift
= 1 + W_TYPE_SIZE
- count
;
2332 /* Shift down the most significant bits. */
2335 /* If we lost any nonzero bits, set the lsb to ensure correct rounding. */
2336 if ((UWtype
)u
<< (W_TYPE_SIZE
- shift
))
2339 /* Convert the one word of data, and rescale. */
2341 if (shift
== W_TYPE_SIZE
)
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;
2349 e
= (Wtype
)1 << shift
;
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) \
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__
2367 #define FUNC __floatundidf
2368 #define FSTYPE DFtype
2369 #define FSSIZE __LIBGCC_DF_MANT_DIG__
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
);
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__))
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
2398 #define REP_BIT ((UDWtype) 1 << (DI_SIZE - FSIZE))
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
2405 if (u
>= ((UDWtype
) 1 << FSIZE
))
2407 if ((UDWtype
) u
& (REP_BIT
- 1))
2409 u
&= ~ (REP_BIT
- 1);
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
);
2421 #if FSSIZE == W_TYPE_SIZE - 1
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
2429 /* If there are no high bits set, fall back to one conversion. */
2431 return (FSTYPE
)(UWtype
)u
;
2433 /* Otherwise, find the power of two. */
2434 UWtype hi
= u
>> W_TYPE_SIZE
;
2436 UWtype count
, shift
;
2437 count_leading_zeros (count
, hi
);
2439 shift
= W_TYPE_SIZE
- count
;
2441 /* Shift down the most significant bits. */
2444 /* If we lost any nonzero bits, set the lsb to ensure correct rounding. */
2445 if ((UWtype
)u
<< (W_TYPE_SIZE
- shift
))
2448 /* Convert the one word of data, and rescale. */
2450 if (shift
== W_TYPE_SIZE
)
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;
2458 e
= (Wtype
)1 << shift
;
2464 #if defined(L_fixunsxfsi) && LIBGCC2_HAS_XF_MODE
2466 __fixunsxfSI (XFtype a
)
2468 if (a
>= - (DFtype
) Wtype_MIN
)
2469 return (Wtype
) (a
+ Wtype_MIN
) - Wtype_MIN
;
2474 #if defined(L_fixunsdfsi) && LIBGCC2_HAS_DF_MODE
2476 __fixunsdfSI (DFtype a
)
2478 if (a
>= - (DFtype
) Wtype_MIN
)
2479 return (Wtype
) (a
+ Wtype_MIN
) - Wtype_MIN
;
2484 #if defined(L_fixunssfsi) && LIBGCC2_HAS_SF_MODE
2486 __fixunssfSI (SFtype a
)
2488 if (a
>= - (SFtype
) Wtype_MIN
)
2489 return (Wtype
) (a
+ Wtype_MIN
) - Wtype_MIN
;
2494 /* Integer power helper used from __builtin_powi for non-constant
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
2518 NAME (TYPE x
, int m
)
2520 unsigned int n
= m
< 0 ? -(unsigned int) m
: (unsigned int) m
;
2521 TYPE y
= n
% 2 ? x
: 1;
2528 return m
< 0 ? 1/y
: y
;
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)
2543 #if defined(L_mulhc3) || defined(L_divhc3)
2544 # define MTYPE HFtype
2545 # define CTYPE HCtype
2546 # define AMTYPE SFtype
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
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
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
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
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__)
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__)
2601 # define RMAX2 (RBIG * RMIN2)
2606 #define CONCAT3(A,B,C) _CONCAT3(A,B,C)
2607 #define _CONCAT3(A,B,C) A##B##C
2609 #define CONCAT2(A,B) _CONCAT2(A,B)
2610 #define _CONCAT2(A,B) A##B
2612 #define isnan(x) __builtin_isnan (x)
2613 #define isfinite(x) __builtin_isfinite (x)
2614 #define isinf(x) __builtin_isinf (x)
2616 #define INFINITY CONCAT2(__builtin_huge_val, CEXT) ()
2619 /* Helpers to make the following code slightly less gross. */
2620 #define COPYSIGN CONCAT2(__builtin_copysign, CEXT)
2621 #define FABS CONCAT2(__builtin_fabs, CEXT)
2623 /* Verify that MTYPE matches up with CEXT. */
2624 extern void *compile_type_assert
[sizeof(INFINITY
) == sizeof(MTYPE
) ? 1 : -1];
2626 /* Ensure that we've lost any extra precision. */
2630 # define TRUNC(x) __asm__ ("" : "=m"(x) : "m"(x))
2633 #if defined(L_mulhc3) || defined(L_mulsc3) || defined(L_muldc3) \
2634 || defined(L_mulxc3) || defined(L_multc3)
2637 CONCAT3(__mul
,MODE
,3) (MTYPE a
, MTYPE b
, MTYPE c
, MTYPE d
)
2639 MTYPE ac
, bd
, ad
, bc
, x
, y
;
2655 if (isnan (x
) && isnan (y
))
2657 /* Recover infinities that computed as NaN + iNaN. */
2659 if (isinf (a
) || isinf (b
))
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
);
2669 if (isinf (c
) || isinf (d
))
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
);
2680 && (isinf (ac
) || isinf (bd
)
2681 || isinf (ad
) || isinf (bc
)))
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
);
2692 x
= INFINITY
* (a
* c
- b
* d
);
2693 y
= INFINITY
* (a
* d
+ b
* c
);
2701 #endif /* complex multiply */
2703 #if defined(L_divhc3) || defined(L_divsc3) || defined(L_divdc3) \
2704 || defined(L_divxc3) || defined(L_divtc3)
2707 CONCAT3(__div
,MODE
,3) (MTYPE a
, MTYPE b
, MTYPE c
, MTYPE d
)
2709 #if defined(L_divhc3) \
2710 || (defined(L_divsc3) && defined(__LIBGCC_HAVE_HWDBL__) )
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. */
2719 AMTYPE aa
, bb
, cc
, dd
;
2728 denom
= (cc
* cc
) + (dd
* dd
);
2729 x
= ((aa
* cc
) + (bb
* dd
)) / denom
;
2730 y
= ((bb
* cc
) - (aa
* dd
)) / denom
;
2733 MTYPE denom
, ratio
, x
, y
;
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.
2742 /* Scale by max(c,d) to reduce chances of denominator overflowing. */
2743 if (FABS (c
) < FABS (d
))
2745 /* Prevent underflow when denominator is near max representable. */
2746 if (FABS (d
) >= RBIG
)
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
)
2765 if (((FABS (a
) < RMIN
) && (FABS (b
) < RMAX2
) && (FABS (d
) < RMAX2
))
2766 || ((FABS (b
) < RMIN
) && (FABS (a
) < RMAX2
)
2767 && (FABS (d
) < RMAX2
)))
2776 denom
= (c
* ratio
) + d
;
2777 /* Choose alternate order of computation if ratio is subnormal. */
2778 if (FABS (ratio
) > RMIN
)
2780 x
= ((a
* ratio
) + b
) / denom
;
2781 y
= ((b
* ratio
) - a
) / denom
;
2785 x
= ((c
* (a
/ d
)) + b
) / denom
;
2786 y
= ((c
* (b
/ d
)) - a
) / denom
;
2791 /* Prevent underflow when denominator is near max representable. */
2792 if (FABS (c
) >= RBIG
)
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
)
2811 if (((FABS (a
) < RMIN
) && (FABS (b
) < RMAX2
) && (FABS (c
) < RMAX2
))
2812 || ((FABS (b
) < RMIN
) && (FABS (a
) < RMAX2
)
2813 && (FABS (c
) < RMAX2
)))
2822 denom
= (d
* ratio
) + c
;
2823 /* Choose alternate order of computation if ratio is subnormal. */
2824 if (FABS (ratio
) > RMIN
)
2826 x
= ((b
* ratio
) + a
) / denom
;
2827 y
= (b
- (a
* ratio
)) / denom
;
2831 x
= (a
+ (d
* (b
/ c
))) / denom
;
2832 y
= (b
- (d
* (a
/ c
))) / denom
;
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
))
2841 if (c
== 0.0 && d
== 0.0 && (!isnan (a
) || !isnan (b
)))
2843 x
= COPYSIGN (INFINITY
, c
) * a
;
2844 y
= COPYSIGN (INFINITY
, c
) * b
;
2846 else if ((isinf (a
) || isinf (b
)) && isfinite (c
) && isfinite (d
))
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
);
2853 else if ((isinf (c
) || isinf (d
)) && isfinite (a
) && isfinite (b
))
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
);
2866 #endif /* complex divide */
2868 #endif /* all complex float routines */
2870 /* From here on down, the routines use normal data types. */
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
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. */
2900 __gcc_bcmp (const unsigned char *s1
, const unsigned char *s2
, size_t size
)
2904 const unsigned char c1
= *s1
++, c2
= *s2
++;
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. */
2919 #ifndef inhibit_libc
2921 #undef NULL /* Avoid errors if stdio.h and our stddef.h mismatch. */
2925 __eprintf (const char *string
, const char *expression
,
2926 unsigned int line
, const char *filename
)
2928 fprintf (stderr
, string
, expression
, line
, filename
);
2937 #ifdef L_clear_cache
2938 /* Clear part of an instruction cache. */
2941 __clear_cache (void *beg
__attribute__((__unused__
)),
2942 void *end
__attribute__((__unused__
)))
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
2948 CLEAR_INSN_CACHE ((char *) beg
, (char *) end
);
2949 #endif /* CLEAR_INSN_CACHE */
2952 #endif /* L_clear_cache */
2956 /* Jump to a trampoline, loading the static chain address. */
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);
2975 mprotect (char *addr
, int len
, int prot
)
2994 if (VirtualProtect (addr
, len
, np
, &op
))
3000 #endif /* WINNT && ! __CYGWIN__ */
3002 #ifdef TRANSFER_FROM_TRAMPOLINE
3003 TRANSFER_FROM_TRAMPOLINE
3005 #endif /* L_trampoline */
3010 #include "gbl-ctors.h"
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. */
3017 #define NAME__MAIN "__main"
3018 #define SYMBOL__MAIN __main
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
3027 #if !defined (HAS_INIT_SECTION) || !defined (OBJECT_FORMAT_ELF)
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. */
3034 #undef __LIBGCC_EH_FRAME_SECTION_NAME__
3037 #ifdef __LIBGCC_EH_FRAME_SECTION_NAME__
3038 #include "unwind-dw2-fde.h"
3039 extern unsigned char __EH_FRAME_BEGIN__
[];
3042 /* Run all the global destructors on exit from the program. */
3045 __do_global_dtors (void)
3047 #ifdef DO_GLOBAL_DTORS_BODY
3048 DO_GLOBAL_DTORS_BODY
;
3050 static func_ptr
*p
= __DTOR_LIST__
+ 1;
3057 #if defined (__LIBGCC_EH_FRAME_SECTION_NAME__) && !defined (HAS_INIT_SECTION)
3059 static int completed
= 0;
3063 __deregister_frame_info (__EH_FRAME_BEGIN__
);
3070 #ifndef HAS_INIT_SECTION
3071 /* Run all the global constructors on entry to the program. */
3074 __do_global_ctors (void)
3076 #ifdef __LIBGCC_EH_FRAME_SECTION_NAME__
3078 static struct object object
;
3079 __register_frame_info (__EH_FRAME_BEGIN__
, &object
);
3082 DO_GLOBAL_CTORS_BODY
;
3083 atexit (__do_global_dtors
);
3085 #endif /* no HAS_INIT_SECTION */
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.
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. */
3096 extern void SYMBOL__MAIN (void);
3100 /* Support recursive calls to `main': run initializers just once. */
3101 static int initialized
;
3105 __do_global_ctors ();
3108 #endif /* no HAS_INIT_SECTION or INVOKE__main */
3110 #endif /* L__main */
3111 #endif /* __CYGWIN__ */
3115 #include "gbl-ctors.h"
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. */
3122 /* We declare the lists here with two elements each,
3123 so that they are valid empty lists if no other definition is loaded.
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.
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};
3135 func_ptr __CTOR_LIST__
[2];
3136 func_ptr __DTOR_LIST__
[2];
3138 #endif /* no __LIBGCC_INIT_SECTION_ASM_OP__ */
3139 #endif /* L_ctors */
3140 #endif /* LIBGCC2_UNITS_PER_WORD <= MIN_UNITS_PER_WORD */