Skip to content

Commit 26bc4d4

Browse files
committed
Optimize ranges::equal for vector<bool>::iterator
1 parent cc45365 commit 26bc4d4

File tree

7 files changed

+441
-221
lines changed

7 files changed

+441
-221
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, move, move_backward}`` algorithms have been optimized for
4747
``std::vector<bool>::iterator``, resulting in a performance improvement of up to 2000x.
4848

49+
- The ``std::ranges::equal`` algorithm has been optimized for ``std::vector<bool>::iterator``, resulting in a performance
50+
improvement of up to 188x.
51+
4952
- Updated formatting library to Unicode 16.0.0.
5053

5154
Deprecations and Removals

libcxx/include/__algorithm/equal.h

Lines changed: 160 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -11,19 +11,27 @@
1111
#define _LIBCPP___ALGORITHM_EQUAL_H
1212

1313
#include <__algorithm/comp.h>
14+
#include <__algorithm/min.h>
1415
#include <__algorithm/unwrap_iter.h>
1516
#include <__config>
1617
#include <__functional/identity.h>
18+
#include <__fwd/bit_reference.h>
1719
#include <__iterator/distance.h>
1820
#include <__iterator/iterator_traits.h>
21+
#include <__memory/pointer_traits.h>
1922
#include <__string/constexpr_c_functions.h>
2023
#include <__type_traits/desugars_to.h>
2124
#include <__type_traits/enable_if.h>
2225
#include <__type_traits/invoke.h>
2326
#include <__type_traits/is_equality_comparable.h>
27+
#include <__type_traits/is_same.h>
2428
#include <__type_traits/is_volatile.h>
2529
#include <__utility/move.h>
2630

31+
#if _LIBCPP_STD_VER >= 20
32+
# include <__functional/ranges_operations.h>
33+
#endif
34+
2735
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
2836
# pragma GCC system_header
2937
#endif
@@ -33,6 +41,136 @@ _LIBCPP_PUSH_MACROS
3341

3442
_LIBCPP_BEGIN_NAMESPACE_STD
3543

44+
template <class _Cp, bool _IsConst1, bool _IsConst2>
45+
[[__nodiscard__]] _LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI bool
46+
__equal_unaligned(__bit_iterator<_Cp, _IsConst1> __first1,
47+
__bit_iterator<_Cp, _IsConst1> __last1,
48+
__bit_iterator<_Cp, _IsConst2> __first2) {
49+
using _It = __bit_iterator<_Cp, _IsConst1>;
50+
using difference_type = typename _It::difference_type;
51+
using __storage_type = typename _It::__storage_type;
52+
53+
const int __bits_per_word = _It::__bits_per_word;
54+
difference_type __n = __last1 - __first1;
55+
if (__n > 0) {
56+
// do first word
57+
if (__first1.__ctz_ != 0) {
58+
unsigned __clz_f = __bits_per_word - __first1.__ctz_;
59+
difference_type __dn = std::min(static_cast<difference_type>(__clz_f), __n);
60+
__n -= __dn;
61+
__storage_type __m = (~__storage_type(0) << __first1.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
62+
__storage_type __b = *__first1.__seg_ & __m;
63+
unsigned __clz_r = __bits_per_word - __first2.__ctz_;
64+
__storage_type __ddn = std::min<__storage_type>(__dn, __clz_r);
65+
__m = (~__storage_type(0) << __first2.__ctz_) & (~__storage_type(0) >> (__clz_r - __ddn));
66+
if (__first2.__ctz_ > __first1.__ctz_) {
67+
if ((*__first2.__seg_ & __m) != (__b << (__first2.__ctz_ - __first1.__ctz_)))
68+
return false;
69+
} else {
70+
if ((*__first2.__seg_ & __m) != (__b >> (__first1.__ctz_ - __first2.__ctz_)))
71+
return false;
72+
}
73+
__first2.__seg_ += (__ddn + __first2.__ctz_) / __bits_per_word;
74+
__first2.__ctz_ = static_cast<unsigned>((__ddn + __first2.__ctz_) % __bits_per_word);
75+
__dn -= __ddn;
76+
if (__dn > 0) {
77+
__m = ~__storage_type(0) >> (__bits_per_word - __dn);
78+
if ((*__first2.__seg_ & __m) != (__b >> (__first1.__ctz_ + __ddn)))
79+
return false;
80+
__first2.__ctz_ = static_cast<unsigned>(__dn);
81+
}
82+
++__first1.__seg_;
83+
// __first1.__ctz_ = 0;
84+
}
85+
// __first1.__ctz_ == 0;
86+
// do middle words
87+
unsigned __clz_r = __bits_per_word - __first2.__ctz_;
88+
__storage_type __m = ~__storage_type(0) << __first2.__ctz_;
89+
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first1.__seg_) {
90+
__storage_type __b = *__first1.__seg_;
91+
if ((*__first2.__seg_ & __m) != (__b << __first2.__ctz_))
92+
return false;
93+
++__first2.__seg_;
94+
if ((*__first2.__seg_ & ~__m) != (__b >> __clz_r))
95+
return false;
96+
}
97+
// do last word
98+
if (__n > 0) {
99+
__m = ~__storage_type(0) >> (__bits_per_word - __n);
100+
__storage_type __b = *__first1.__seg_ & __m;
101+
__storage_type __dn = std::min(__n, static_cast<difference_type>(__clz_r));
102+
__m = (~__storage_type(0) << __first2.__ctz_) & (~__storage_type(0) >> (__clz_r - __dn));
103+
if ((*__first2.__seg_ & __m) != (__b << __first2.__ctz_))
104+
return false;
105+
__first2.__seg_ += (__dn + __first2.__ctz_) / __bits_per_word;
106+
__first2.__ctz_ = static_cast<unsigned>((__dn + __first2.__ctz_) % __bits_per_word);
107+
__n -= __dn;
108+
if (__n > 0) {
109+
__m = ~__storage_type(0) >> (__bits_per_word - __n);
110+
if ((*__first2.__seg_ & __m) != (__b >> __dn))
111+
return false;
112+
}
113+
}
114+
}
115+
return true;
116+
}
117+
118+
template <class _Cp, bool _IsConst1, bool _IsConst2>
119+
[[__nodiscard__]] _LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI bool
120+
__equal_aligned(__bit_iterator<_Cp, _IsConst1> __first1,
121+
__bit_iterator<_Cp, _IsConst1> __last1,
122+
__bit_iterator<_Cp, _IsConst2> __first2) {
123+
using _It = __bit_iterator<_Cp, _IsConst1>;
124+
using difference_type = typename _It::difference_type;
125+
using __storage_type = typename _It::__storage_type;
126+
127+
const int __bits_per_word = _It::__bits_per_word;
128+
difference_type __n = __last1 - __first1;
129+
if (__n > 0) {
130+
// do first word
131+
if (__first1.__ctz_ != 0) {
132+
unsigned __clz = __bits_per_word - __first1.__ctz_;
133+
difference_type __dn = std::min(static_cast<difference_type>(__clz), __n);
134+
__n -= __dn;
135+
__storage_type __m = (~__storage_type(0) << __first1.__ctz_) & (~__storage_type(0) >> (__clz - __dn));
136+
if ((*__first2.__seg_ & __m) != (*__first1.__seg_ & __m))
137+
return false;
138+
++__first2.__seg_;
139+
++__first1.__seg_;
140+
// __first1.__ctz_ = 0;
141+
// __first2.__ctz_ = 0;
142+
}
143+
// __first1.__ctz_ == 0;
144+
// __first2.__ctz_ == 0;
145+
// do middle words
146+
for (; __n >= __bits_per_word; __n -= __bits_per_word, ++__first1.__seg_, ++__first2.__seg_)
147+
if (*__first2.__seg_ != *__first1.__seg_)
148+
return false;
149+
// do last word
150+
if (__n > 0) {
151+
__storage_type __m = ~__storage_type(0) >> (__bits_per_word - __n);
152+
if ((*__first2.__seg_ & __m) != (*__first1.__seg_ & __m))
153+
return false;
154+
}
155+
}
156+
return true;
157+
}
158+
159+
template <class _Cp,
160+
bool _IsConst1,
161+
bool _IsConst2,
162+
class _BinaryPredicate,
163+
__enable_if_t<__desugars_to_v<__equal_tag, _BinaryPredicate, bool, bool>, int> = 0>
164+
[[__nodiscard__]] inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 bool __equal_iter_impl(
165+
__bit_iterator<_Cp, _IsConst1> __first1,
166+
__bit_iterator<_Cp, _IsConst1> __last1,
167+
__bit_iterator<_Cp, _IsConst2> __first2,
168+
_BinaryPredicate) {
169+
if (__first1.__ctz_ == __first2.__ctz_)
170+
return std::__equal_aligned(__first1, __last1, __first2);
171+
return std::__equal_unaligned(__first1, __last1, __first2);
172+
}
173+
36174
template <class _InputIterator1, class _InputIterator2, class _BinaryPredicate>
37175
[[__nodiscard__]] inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 bool __equal_iter_impl(
38176
_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _BinaryPredicate& __pred) {
@@ -94,6 +232,28 @@ __equal_impl(_Tp* __first1, _Tp* __last1, _Up* __first2, _Up*, _Pred&, _Proj1&,
94232
return std::__constexpr_memcmp_equal(__first1, __first2, __element_count(__last1 - __first1));
95233
}
96234

235+
template <class _Cp,
236+
bool _IsConst1,
237+
bool _IsConst2,
238+
class _Pred,
239+
class _Proj1,
240+
class _Proj2,
241+
__enable_if_t<__desugars_to_v<__equal_tag, _Pred, bool, bool> && __is_identity<_Proj1>::value &&
242+
__is_identity<_Proj2>::value,
243+
int> = 0>
244+
[[__nodiscard__]] inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 bool __equal_impl(
245+
__bit_iterator<_Cp, _IsConst1> __first1,
246+
__bit_iterator<_Cp, _IsConst1> __last1,
247+
__bit_iterator<_Cp, _IsConst2> __first2,
248+
__bit_iterator<_Cp, _IsConst2>,
249+
_Pred&,
250+
_Proj1&,
251+
_Proj2&) {
252+
if (__first1.__ctz_ == __first2.__ctz_)
253+
return std::__equal_aligned(__first1, __last1, __first2);
254+
return std::__equal_unaligned(__first1, __last1, __first2);
255+
}
256+
97257
template <class _InputIterator1, class _InputIterator2, class _BinaryPredicate>
98258
[[__nodiscard__]] inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 bool
99259
equal(_InputIterator1 __first1,

0 commit comments

Comments
 (0)