Skip to content

Commit 0bfcf34

Browse files
winner245winner245
committed
Optimize ranges::swap_ranges for vector<bool>::iterator
1 parent c6c6475 commit 0bfcf34

File tree

6 files changed

+416
-296
lines changed

6 files changed

+416
-296
lines changed

‎libcxx/include/__algorithm/swap_ranges.h‎

Lines changed: 162 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,9 +10,12 @@
1010
#define _LIBCPP___ALGORITHM_SWAP_RANGES_H
1111

1212
#include <__algorithm/iterator_operations.h>
13+
#include <__algorithm/min.h>
1314
#include <__config>
15+
#include <__fwd/bit_reference.h>
1416
#include <__utility/move.h>
1517
#include <__utility/pair.h>
18+
#include <__utility/swap.h>
1619

1720
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
1821
# pragma GCC system_header
@@ -23,6 +26,165 @@ _LIBCPP_PUSH_MACROS
2326

2427
_LIBCPP_BEGIN_NAMESPACE_STD
2528

29+
template <class _Cl, class _Cr>
30+
_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 __bit_iterator<_Cr, false> __swap_ranges_aligned(
31+
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
32+
using _I1 = __bit_iterator<_Cl, false>;
33+
using difference_type = typename _I1::difference_type;
34+
using __storage_type = typename _I1::__storage_type;
35+
36+
const int __bits_per_word = _I1::__bits_per_word;
37+
difference_type __n = __last - __first;
38+
if (__n > 0) {
39+
// do first word
40+
if (__first.__ctz_ != 0) {
41+
unsigned __clz = __bits_per_word - __first.__ctz_;
42+
difference_type __dn = std::min(static_cast<difference_type>(__clz), __n);
43+
__n -= __dn;
44+
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
45+
__storage_type __b1 = *__first.__seg_ & __m;
46+
*__first.__seg_ &= ~__m;
47+
__storage_type __b2 = *__result.__seg_ & __m;
48+
*__result.__seg_ &= ~__m;
49+
*__result.__seg_ |= __b1;
50+
*__first.__seg_ |= __b2;
51+
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
52+
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
53+
++__first.__seg_;
54+
// __first.__ctz_ = 0;
55+
}
56+
// __first.__ctz_ == 0;
57+
// do middle words
58+
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_, ++__result.__seg_)
59+
swap(*__first.__seg_, *__result.__seg_);
60+
// do last word
61+
if (__n > 0) {
62+
__storage_type __m = ~__storage_type(0) >> (__bits_per_word - __n);
63+
__storage_type __b1 = *__first.__seg_ & __m;
64+
*__first.__seg_ &= ~__m;
65+
__storage_type __b2 = *__result.__seg_ & __m;
66+
*__result.__seg_ &= ~__m;
67+
*__result.__seg_ |= __b1;
68+
*__first.__seg_ |= __b2;
69+
__result.__ctz_ = static_cast<unsigned>(__n);
70+
}
71+
}
72+
return __result;
73+
}
74+
75+
template <class _Cl, class _Cr>
76+
_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 __bit_iterator<_Cr, false> __swap_ranges_unaligned(
77+
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
78+
using _I1 = __bit_iterator<_Cl, false>;
79+
using difference_type = typename _I1::difference_type;
80+
using __storage_type = typename _I1::__storage_type;
81+
82+
const int __bits_per_word = _I1::__bits_per_word;
83+
difference_type __n = __last - __first;
84+
if (__n > 0) {
85+
// do first word
86+
if (__first.__ctz_ != 0) {
87+
unsigned __clz_f = __bits_per_word - __first.__ctz_;
88+
difference_type __dn = std::min(static_cast<difference_type>(__clz_f), __n);
89+
__n -= __dn;
90+
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
91+
__storage_type __b1 = *__first.__seg_ & __m;
92+
*__first.__seg_ &= ~__m;
93+
unsigned __clz_r = __bits_per_word - __result.__ctz_;
94+
__storage_type __ddn = std::min<__storage_type>(__dn, __clz_r);
95+
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __ddn));
96+
__storage_type __b2 = *__result.__seg_ & __m;
97+
*__result.__seg_ &= ~__m;
98+
if (__result.__ctz_ > __first.__ctz_) {
99+
unsigned __s = __result.__ctz_ - __first.__ctz_;
100+
*__result.__seg_ |= __b1 << __s;
101+
*__first.__seg_ |= __b2 >> __s;
102+
} else {
103+
unsigned __s = __first.__ctz_ - __result.__ctz_;
104+
*__result.__seg_ |= __b1 >> __s;
105+
*__first.__seg_ |= __b2 << __s;
106+
}
107+
__result.__seg_ += (__ddn + __result.__ctz_) / __bits_per_word;
108+
__result.__ctz_ = static_cast<unsigned>((__ddn + __result.__ctz_) % __bits_per_word);
109+
__dn -= __ddn;
110+
if (__dn > 0) {
111+
__m = ~__storage_type(0) >> (__bits_per_word - __dn);
112+
__b2 = *__result.__seg_ & __m;
113+
*__result.__seg_ &= ~__m;
114+
unsigned __s = __first.__ctz_ + __ddn;
115+
*__result.__seg_ |= __b1 >> __s;
116+
*__first.__seg_ |= __b2 << __s;
117+
__result.__ctz_ = static_cast<unsigned>(__dn);
118+
}
119+
++__first.__seg_;
120+
// __first.__ctz_ = 0;
121+
}
122+
// __first.__ctz_ == 0;
123+
// do middle words
124+
__storage_type __m = ~__storage_type(0) << __result.__ctz_;
125+
unsigned __clz_r = __bits_per_word - __result.__ctz_;
126+
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_) {
127+
__storage_type __b1 = *__first.__seg_;
128+
__storage_type __b2 = *__result.__seg_ & __m;
129+
*__result.__seg_ &= ~__m;
130+
*__result.__seg_ |= __b1 << __result.__ctz_;
131+
*__first.__seg_ = __b2 >> __result.__ctz_;
132+
++__result.__seg_;
133+
__b2 = *__result.__seg_ & ~__m;
134+
*__result.__seg_ &= __m;
135+
*__result.__seg_ |= __b1 >> __clz_r;
136+
*__first.__seg_ |= __b2 << __clz_r;
137+
}
138+
// do last word
139+
if (__n > 0) {
140+
__m = ~__storage_type(0) >> (__bits_per_word - __n);
141+
__storage_type __b1 = *__first.__seg_ & __m;
142+
*__first.__seg_ &= ~__m;
143+
__storage_type __dn = std::min<__storage_type>(__n, __clz_r);
144+
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __dn));
145+
__storage_type __b2 = *__result.__seg_ & __m;
146+
*__result.__seg_ &= ~__m;
147+
*__result.__seg_ |= __b1 << __result.__ctz_;
148+
*__first.__seg_ |= __b2 >> __result.__ctz_;
149+
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
150+
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
151+
__n -= __dn;
152+
if (__n > 0) {
153+
__m = ~__storage_type(0) >> (__bits_per_word - __n);
154+
__b2 = *__result.__seg_ & __m;
155+
*__result.__seg_ &= ~__m;
156+
*__result.__seg_ |= __b1 >> __dn;
157+
*__first.__seg_ |= __b2 << __dn;
158+
__result.__ctz_ = static_cast<unsigned>(__n);
159+
}
160+
}
161+
}
162+
return __result;
163+
}
164+
165+
// 2+1 iterators: size2 >= size1; used by std::swap_ranges.
166+
template <class, class _Cl, class _Cr>
167+
_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<__bit_iterator<_Cl, false>, __bit_iterator<_Cr, false> >
168+
__swap_ranges(__bit_iterator<_Cl, false> __first1,
169+
__bit_iterator<_Cl, false> __last1,
170+
__bit_iterator<_Cr, false> __first2) {
171+
if (__first1.__ctz_ == __first2.__ctz_)
172+
return std::make_pair(__last1, std::__swap_ranges_aligned(__first1, __last1, __first2));
173+
return std::make_pair(__last1, std::__swap_ranges_unaligned(__first1, __last1, __first2));
174+
}
175+
176+
// 2+2 iterators: used by std::ranges::swap_ranges.
177+
template <class _AlgPolicy, class _Cl, class _Cr>
178+
_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<__bit_iterator<_Cl, false>, __bit_iterator<_Cr, false> >
179+
__swap_ranges(__bit_iterator<_Cl, false> __first1,
180+
__bit_iterator<_Cl, false> __last1,
181+
__bit_iterator<_Cr, false> __first2,
182+
__bit_iterator<_Cr, false> __last2) {
183+
if (__last1 - __first1 < __last2 - __first2)
184+
return std::make_pair(__last1, std::__swap_ranges<_AlgPolicy>(__first1, __last1, __first2).second);
185+
return std::make_pair(std::__swap_ranges<_AlgPolicy>(__first2, __last2, __first1).second, __last2);
186+
}
187+
26188
// 2+2 iterators: the shorter size will be used.
27189
template <class _AlgPolicy, class _ForwardIterator1, class _Sentinel1, class _ForwardIterator2, class _Sentinel2>
28190
_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_ForwardIterator1, _ForwardIterator2>

‎libcxx/include/__bit_reference‎

Lines changed: 7 additions & 151 deletions
Original file line numberDiff line numberDiff line change
@@ -12,6 +12,7 @@
1212

1313
#include <__algorithm/copy_n.h>
1414
#include <__algorithm/min.h>
15+
#include <__algorithm/swap_ranges.h>
1516
#include <__bit/countr.h>
1617
#include <__compare/ordering.h>
1718
#include <__config>
@@ -24,6 +25,7 @@
2425
#include <__type_traits/conditional.h>
2526
#include <__type_traits/is_constant_evaluated.h>
2627
#include <__type_traits/void_t.h>
28+
#include <__utility/pair.h>
2729
#include <__utility/swap.h>
2830

2931
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
@@ -451,152 +453,6 @@ inline _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> move_backward(
451453
return std::copy_backward(__first, __last, __result);
452454
}
453455

454-
// swap_ranges
455-
456-
template <class _Cl, class _Cr>
457-
_LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> __swap_ranges_aligned(
458-
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
459-
using _I1 = __bit_iterator<_Cl, false>;
460-
using difference_type = typename _I1::difference_type;
461-
using __storage_type = typename _I1::__storage_type;
462-
463-
const int __bits_per_word = _I1::__bits_per_word;
464-
difference_type __n = __last - __first;
465-
if (__n > 0) {
466-
// do first word
467-
if (__first.__ctz_ != 0) {
468-
unsigned __clz = __bits_per_word - __first.__ctz_;
469-
difference_type __dn = std::min(static_cast<difference_type>(__clz), __n);
470-
__n -= __dn;
471-
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
472-
__storage_type __b1 = *__first.__seg_ & __m;
473-
*__first.__seg_ &= ~__m;
474-
__storage_type __b2 = *__result.__seg_ & __m;
475-
*__result.__seg_ &= ~__m;
476-
*__result.__seg_ |= __b1;
477-
*__first.__seg_ |= __b2;
478-
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
479-
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
480-
++__first.__seg_;
481-
// __first.__ctz_ = 0;
482-
}
483-
// __first.__ctz_ == 0;
484-
// do middle words
485-
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_, ++__result.__seg_)
486-
swap(*__first.__seg_, *__result.__seg_);
487-
// do last word
488-
if (__n > 0) {
489-
__storage_type __m = ~__storage_type(0) >> (__bits_per_word - __n);
490-
__storage_type __b1 = *__first.__seg_ & __m;
491-
*__first.__seg_ &= ~__m;
492-
__storage_type __b2 = *__result.__seg_ & __m;
493-
*__result.__seg_ &= ~__m;
494-
*__result.__seg_ |= __b1;
495-
*__first.__seg_ |= __b2;
496-
__result.__ctz_ = static_cast<unsigned>(__n);
497-
}
498-
}
499-
return __result;
500-
}
501-
502-
template <class _Cl, class _Cr>
503-
_LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> __swap_ranges_unaligned(
504-
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
505-
using _I1 = __bit_iterator<_Cl, false>;
506-
using difference_type = typename _I1::difference_type;
507-
using __storage_type = typename _I1::__storage_type;
508-
509-
const int __bits_per_word = _I1::__bits_per_word;
510-
difference_type __n = __last - __first;
511-
if (__n > 0) {
512-
// do first word
513-
if (__first.__ctz_ != 0) {
514-
unsigned __clz_f = __bits_per_word - __first.__ctz_;
515-
difference_type __dn = std::min(static_cast<difference_type>(__clz_f), __n);
516-
__n -= __dn;
517-
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
518-
__storage_type __b1 = *__first.__seg_ & __m;
519-
*__first.__seg_ &= ~__m;
520-
unsigned __clz_r = __bits_per_word - __result.__ctz_;
521-
__storage_type __ddn = std::min<__storage_type>(__dn, __clz_r);
522-
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __ddn));
523-
__storage_type __b2 = *__result.__seg_ & __m;
524-
*__result.__seg_ &= ~__m;
525-
if (__result.__ctz_ > __first.__ctz_) {
526-
unsigned __s = __result.__ctz_ - __first.__ctz_;
527-
*__result.__seg_ |= __b1 << __s;
528-
*__first.__seg_ |= __b2 >> __s;
529-
} else {
530-
unsigned __s = __first.__ctz_ - __result.__ctz_;
531-
*__result.__seg_ |= __b1 >> __s;
532-
*__first.__seg_ |= __b2 << __s;
533-
}
534-
__result.__seg_ += (__ddn + __result.__ctz_) / __bits_per_word;
535-
__result.__ctz_ = static_cast<unsigned>((__ddn + __result.__ctz_) % __bits_per_word);
536-
__dn -= __ddn;
537-
if (__dn > 0) {
538-
__m = ~__storage_type(0) >> (__bits_per_word - __dn);
539-
__b2 = *__result.__seg_ & __m;
540-
*__result.__seg_ &= ~__m;
541-
unsigned __s = __first.__ctz_ + __ddn;
542-
*__result.__seg_ |= __b1 >> __s;
543-
*__first.__seg_ |= __b2 << __s;
544-
__result.__ctz_ = static_cast<unsigned>(__dn);
545-
}
546-
++__first.__seg_;
547-
// __first.__ctz_ = 0;
548-
}
549-
// __first.__ctz_ == 0;
550-
// do middle words
551-
__storage_type __m = ~__storage_type(0) << __result.__ctz_;
552-
unsigned __clz_r = __bits_per_word - __result.__ctz_;
553-
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_) {
554-
__storage_type __b1 = *__first.__seg_;
555-
__storage_type __b2 = *__result.__seg_ & __m;
556-
*__result.__seg_ &= ~__m;
557-
*__result.__seg_ |= __b1 << __result.__ctz_;
558-
*__first.__seg_ = __b2 >> __result.__ctz_;
559-
++__result.__seg_;
560-
__b2 = *__result.__seg_ & ~__m;
561-
*__result.__seg_ &= __m;
562-
*__result.__seg_ |= __b1 >> __clz_r;
563-
*__first.__seg_ |= __b2 << __clz_r;
564-
}
565-
// do last word
566-
if (__n > 0) {
567-
__m = ~__storage_type(0) >> (__bits_per_word - __n);
568-
__storage_type __b1 = *__first.__seg_ & __m;
569-
*__first.__seg_ &= ~__m;
570-
__storage_type __dn = std::min<__storage_type>(__n, __clz_r);
571-
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __dn));
572-
__storage_type __b2 = *__result.__seg_ & __m;
573-
*__result.__seg_ &= ~__m;
574-
*__result.__seg_ |= __b1 << __result.__ctz_;
575-
*__first.__seg_ |= __b2 >> __result.__ctz_;
576-
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
577-
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
578-
__n -= __dn;
579-
if (__n > 0) {
580-
__m = ~__storage_type(0) >> (__bits_per_word - __n);
581-
__b2 = *__result.__seg_ & __m;
582-
*__result.__seg_ &= ~__m;
583-
*__result.__seg_ |= __b1 >> __dn;
584-
*__first.__seg_ |= __b2 << __dn;
585-
__result.__ctz_ = static_cast<unsigned>(__n);
586-
}
587-
}
588-
}
589-
return __result;
590-
}
591-
592-
template <class _Cl, class _Cr>
593-
inline _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> swap_ranges(
594-
__bit_iterator<_Cl, false> __first1, __bit_iterator<_Cl, false> __last1, __bit_iterator<_Cr, false> __first2) {
595-
if (__first1.__ctz_ == __first2.__ctz_)
596-
return std::__swap_ranges_aligned(__first1, __last1, __first2);
597-
return std::__swap_ranges_unaligned(__first1, __last1, __first2);
598-
}
599-
600456
// rotate
601457

602458
template <class _Cp>
@@ -1001,14 +857,14 @@ private:
1001857
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Dp, false>
1002858
copy_backward(__bit_iterator<_Dp, _IC> __first, __bit_iterator<_Dp, _IC> __last, __bit_iterator<_Dp, false> __result);
1003859
template <class _Cl, class _Cr>
1004-
friend __bit_iterator<_Cr, false>
860+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Cr, false>
1005861
__swap_ranges_aligned(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
1006862
template <class _Cl, class _Cr>
1007-
friend __bit_iterator<_Cr, false>
863+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Cr, false>
1008864
__swap_ranges_unaligned(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
1009-
template <class _Cl, class _Cr>
1010-
friend __bit_iterator<_Cr, false>
1011-
swap_ranges(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
865+
template <class, class _Cl, class _Cr>
866+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend pair<__bit_iterator<_Cl, false>, __bit_iterator<_Cr, false> >
867+
__swap_ranges(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
1012868
template <class _Dp>
1013869
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Dp, false>
1014870
rotate(__bit_iterator<_Dp, false>, __bit_iterator<_Dp, false>, __bit_iterator<_Dp, false>);

0 commit comments

Comments
 (0)