Skip to content

Commit 3a220ac

Browse files
winner245winner245
committed
Optimize ranges::swap_ranges for vector<bool>::iterator
1 parent 4590f75 commit 3a220ac

File tree

7 files changed

+422
-296
lines changed

7 files changed

+422
-296
lines changed

‎libcxx/docs/ReleaseNotes/21.rst‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -46,6 +46,9 @@ Improvements and New Features
4646
- The ``std::ranges::{copy, copy_n, copy_backward}`` algorithms have been optimized for ``std::vector<bool>::iterator``\s,
4747
resulting in a performance improvement of up to 2000x.
4848

49+
- The ``std::ranges::swap_ranges`` algorithm has been optimized for ``std::vector<bool>::iterator``, resulting in a
50+
performance improvement of up to 611x.
51+
4952
- Updated formatting library to Unicode 16.0.0.
5053

5154
Deprecations and Removals

‎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: 6 additions & 151 deletions
Original file line numberDiff line numberDiff line change
@@ -14,6 +14,7 @@
1414
#include <__algorithm/copy_backward.h>
1515
#include <__algorithm/copy_n.h>
1616
#include <__algorithm/min.h>
17+
#include <__algorithm/swap_ranges.h>
1718
#include <__assert>
1819
#include <__bit/countr.h>
1920
#include <__compare/ordering.h>
@@ -226,152 +227,6 @@ inline _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cp, false> move_backward(
226227
return std::copy_backward(__first, __last, __result);
227228
}
228229

229-
// swap_ranges
230-
231-
template <class _Cl, class _Cr>
232-
_LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> __swap_ranges_aligned(
233-
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
234-
using _I1 = __bit_iterator<_Cl, false>;
235-
using difference_type = typename _I1::difference_type;
236-
using __storage_type = typename _I1::__storage_type;
237-
238-
const int __bits_per_word = _I1::__bits_per_word;
239-
difference_type __n = __last - __first;
240-
if (__n > 0) {
241-
// do first word
242-
if (__first.__ctz_ != 0) {
243-
unsigned __clz = __bits_per_word - __first.__ctz_;
244-
difference_type __dn = std::min(static_cast<difference_type>(__clz), __n);
245-
__n -= __dn;
246-
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
247-
__storage_type __b1 = *__first.__seg_ & __m;
248-
*__first.__seg_ &= ~__m;
249-
__storage_type __b2 = *__result.__seg_ & __m;
250-
*__result.__seg_ &= ~__m;
251-
*__result.__seg_ |= __b1;
252-
*__first.__seg_ |= __b2;
253-
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
254-
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
255-
++__first.__seg_;
256-
// __first.__ctz_ = 0;
257-
}
258-
// __first.__ctz_ == 0;
259-
// do middle words
260-
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_, ++__result.__seg_)
261-
swap(*__first.__seg_, *__result.__seg_);
262-
// do last word
263-
if (__n > 0) {
264-
__storage_type __m = ~__storage_type(0) >> (__bits_per_word - __n);
265-
__storage_type __b1 = *__first.__seg_ & __m;
266-
*__first.__seg_ &= ~__m;
267-
__storage_type __b2 = *__result.__seg_ & __m;
268-
*__result.__seg_ &= ~__m;
269-
*__result.__seg_ |= __b1;
270-
*__first.__seg_ |= __b2;
271-
__result.__ctz_ = static_cast<unsigned>(__n);
272-
}
273-
}
274-
return __result;
275-
}
276-
277-
template <class _Cl, class _Cr>
278-
_LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> __swap_ranges_unaligned(
279-
__bit_iterator<_Cl, false> __first, __bit_iterator<_Cl, false> __last, __bit_iterator<_Cr, false> __result) {
280-
using _I1 = __bit_iterator<_Cl, false>;
281-
using difference_type = typename _I1::difference_type;
282-
using __storage_type = typename _I1::__storage_type;
283-
284-
const int __bits_per_word = _I1::__bits_per_word;
285-
difference_type __n = __last - __first;
286-
if (__n > 0) {
287-
// do first word
288-
if (__first.__ctz_ != 0) {
289-
unsigned __clz_f = __bits_per_word - __first.__ctz_;
290-
difference_type __dn = std::min(static_cast<difference_type>(__clz_f), __n);
291-
__n -= __dn;
292-
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
293-
__storage_type __b1 = *__first.__seg_ & __m;
294-
*__first.__seg_ &= ~__m;
295-
unsigned __clz_r = __bits_per_word - __result.__ctz_;
296-
__storage_type __ddn = std::min<__storage_type>(__dn, __clz_r);
297-
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __ddn));
298-
__storage_type __b2 = *__result.__seg_ & __m;
299-
*__result.__seg_ &= ~__m;
300-
if (__result.__ctz_ > __first.__ctz_) {
301-
unsigned __s = __result.__ctz_ - __first.__ctz_;
302-
*__result.__seg_ |= __b1 << __s;
303-
*__first.__seg_ |= __b2 >> __s;
304-
} else {
305-
unsigned __s = __first.__ctz_ - __result.__ctz_;
306-
*__result.__seg_ |= __b1 >> __s;
307-
*__first.__seg_ |= __b2 << __s;
308-
}
309-
__result.__seg_ += (__ddn + __result.__ctz_) / __bits_per_word;
310-
__result.__ctz_ = static_cast<unsigned>((__ddn + __result.__ctz_) % __bits_per_word);
311-
__dn -= __ddn;
312-
if (__dn > 0) {
313-
__m = ~__storage_type(0) >> (__bits_per_word - __dn);
314-
__b2 = *__result.__seg_ & __m;
315-
*__result.__seg_ &= ~__m;
316-
unsigned __s = __first.__ctz_ + __ddn;
317-
*__result.__seg_ |= __b1 >> __s;
318-
*__first.__seg_ |= __b2 << __s;
319-
__result.__ctz_ = static_cast<unsigned>(__dn);
320-
}
321-
++__first.__seg_;
322-
// __first.__ctz_ = 0;
323-
}
324-
// __first.__ctz_ == 0;
325-
// do middle words
326-
__storage_type __m = ~__storage_type(0) << __result.__ctz_;
327-
unsigned __clz_r = __bits_per_word - __result.__ctz_;
328-
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first.__seg_) {
329-
__storage_type __b1 = *__first.__seg_;
330-
__storage_type __b2 = *__result.__seg_ & __m;
331-
*__result.__seg_ &= ~__m;
332-
*__result.__seg_ |= __b1 << __result.__ctz_;
333-
*__first.__seg_ = __b2 >> __result.__ctz_;
334-
++__result.__seg_;
335-
__b2 = *__result.__seg_ & ~__m;
336-
*__result.__seg_ &= __m;
337-
*__result.__seg_ |= __b1 >> __clz_r;
338-
*__first.__seg_ |= __b2 << __clz_r;
339-
}
340-
// do last word
341-
if (__n > 0) {
342-
__m = ~__storage_type(0) >> (__bits_per_word - __n);
343-
__storage_type __b1 = *__first.__seg_ & __m;
344-
*__first.__seg_ &= ~__m;
345-
__storage_type __dn = std::min<__storage_type>(__n, __clz_r);
346-
__m = (~__storage_type(0) << __result.__ctz_) & (~__storage_type(0) >> (__clz_r - __dn));
347-
__storage_type __b2 = *__result.__seg_ & __m;
348-
*__result.__seg_ &= ~__m;
349-
*__result.__seg_ |= __b1 << __result.__ctz_;
350-
*__first.__seg_ |= __b2 >> __result.__ctz_;
351-
__result.__seg_ += (__dn + __result.__ctz_) / __bits_per_word;
352-
__result.__ctz_ = static_cast<unsigned>((__dn + __result.__ctz_) % __bits_per_word);
353-
__n -= __dn;
354-
if (__n > 0) {
355-
__m = ~__storage_type(0) >> (__bits_per_word - __n);
356-
__b2 = *__result.__seg_ & __m;
357-
*__result.__seg_ &= ~__m;
358-
*__result.__seg_ |= __b1 >> __dn;
359-
*__first.__seg_ |= __b2 << __dn;
360-
__result.__ctz_ = static_cast<unsigned>(__n);
361-
}
362-
}
363-
}
364-
return __result;
365-
}
366-
367-
template <class _Cl, class _Cr>
368-
inline _LIBCPP_HIDE_FROM_ABI __bit_iterator<_Cr, false> swap_ranges(
369-
__bit_iterator<_Cl, false> __first1, __bit_iterator<_Cl, false> __last1, __bit_iterator<_Cr, false> __first2) {
370-
if (__first1.__ctz_ == __first2.__ctz_)
371-
return std::__swap_ranges_aligned(__first1, __last1, __first2);
372-
return std::__swap_ranges_unaligned(__first1, __last1, __first2);
373-
}
374-
375230
// rotate
376231

377232
template <class _Cp>
@@ -776,14 +631,14 @@ private:
776631
template <class _AlgPolicy>
777632
friend struct __copy_backward_impl;
778633
template <class _Cl, class _Cr>
779-
friend __bit_iterator<_Cr, false>
634+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Cr, false>
780635
__swap_ranges_aligned(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
781636
template <class _Cl, class _Cr>
782-
friend __bit_iterator<_Cr, false>
637+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Cr, false>
783638
__swap_ranges_unaligned(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
784-
template <class _Cl, class _Cr>
785-
friend __bit_iterator<_Cr, false>
786-
swap_ranges(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
639+
template <class, class _Cl, class _Cr>
640+
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend pair<__bit_iterator<_Cl, false>, __bit_iterator<_Cr, false> >
641+
__swap_ranges(__bit_iterator<_Cl, false>, __bit_iterator<_Cl, false>, __bit_iterator<_Cr, false>);
787642
template <class _Dp>
788643
_LIBCPP_CONSTEXPR_SINCE_CXX20 friend __bit_iterator<_Dp, false>
789644
rotate(__bit_iterator<_Dp, false>, __bit_iterator<_Dp, false>, __bit_iterator<_Dp, false>);

0 commit comments

Comments
 (0)