Peng Liu d0cee6939a
[libc++] Optimize std::{,ranges}::{fill,fill_n} for segmented iterators (#132665)
This patch optimizes `std::fill`, `std::fill_n`, `std::ranges::fill`,
and `std::ranges::fill_n` for segmented iterators, achieving substantial
performance improvements. Specifically, for `deque<int>` iterators, the
performance improvements are above 10x for all these algorithms. The
optimization also enables filling segmented memory of `deque<int>` to
approach the performance of filling contiguous memory of `vector<int>`.


Benchmark results comparing the before and after implementations are
provided below. For additional context, we’ve included `vector<int>`
results, which remain unchanged, as this patch specifically targets
segmented iterators and leaves non-segmented iterator behavior
untouched.



Fixes two subtasks outlined in #102817.

#### `fill_n`

```
-----------------------------------------------------------------------------
Benchmark                                Before            After      Speedup
-----------------------------------------------------------------------------
std::fill_n(deque<int>)/32              11.4 ns          2.28 ns        5.0x
std::fill_n(deque<int>)/50              19.7 ns          3.40 ns        5.8x
std::fill_n(deque<int>)/1024             391 ns          37.3 ns       10.5x
std::fill_n(deque<int>)/8192            3174 ns           301 ns       10.5x
std::fill_n(deque<int>)/65536          26504 ns          2951 ns        9.0x
std::fill_n(deque<int>)/1048576       407960 ns         80658 ns        5.1x
rng::fill_n(deque<int>)/32              14.3 ns          2.15 ns        6.6x
rng::fill_n(deque<int>)/50              20.2 ns          3.22 ns        6.3x
rng::fill_n(deque<int>)/1024             381 ns          37.8 ns       10.1x
rng::fill_n(deque<int>)/8192            3101 ns           294 ns       10.5x
rng::fill_n(deque<int>)/65536          25098 ns          2926 ns        8.6x
rng::fill_n(deque<int>)/1048576       394342 ns         78874 ns        5.0x
std::fill_n(vector<int>)/32             1.76 ns          1.72 ns        1.0x
std::fill_n(vector<int>)/50             3.00 ns          2.73 ns        1.1x
std::fill_n(vector<int>)/1024           38.4 ns          37.9 ns        1.0x
std::fill_n(vector<int>)/8192            258 ns           252 ns        1.0x
std::fill_n(vector<int>)/65536          2993 ns          2889 ns        1.0x
std::fill_n(vector<int>)/1048576       80328 ns         80468 ns        1.0x
rng::fill_n(vector<int>)/32             1.99 ns          1.35 ns        1.5x
rng::fill_n(vector<int>)/50             2.66 ns          2.12 ns        1.3x
rng::fill_n(vector<int>)/1024           37.7 ns          35.8 ns        1.1x
rng::fill_n(vector<int>)/8192            253 ns           250 ns        1.0x
rng::fill_n(vector<int>)/65536          2922 ns          2930 ns        1.0x
rng::fill_n(vector<int>)/1048576       79739 ns         79742 ns        1.0x
```

#### `fill`

```
--------------------------------------------------------------------------
Benchmark                              Before            After     Speedup
--------------------------------------------------------------------------
std::fill(deque<int>)/32              13.7 ns          2.45 ns        5.6x
std::fill(deque<int>)/50              21.7 ns          4.57 ns        4.7x
std::fill(deque<int>)/1024             367 ns          38.5 ns        9.5x
std::fill(deque<int>)/8192            2896 ns           247 ns       11.7x
std::fill(deque<int>)/65536          23723 ns          2907 ns        8.2x
std::fill(deque<int>)/1048576       379043 ns         79885 ns        4.7x
rng::fill(deque<int>)/32              13.6 ns          2.70 ns        5.0x
rng::fill(deque<int>)/50              23.4 ns          3.94 ns        5.9x
rng::fill(deque<int>)/1024             377 ns          37.9 ns        9.9x
rng::fill(deque<int>)/8192            2914 ns           286 ns       10.2x
rng::fill(deque<int>)/65536          23612 ns          2939 ns        8.0x
rng::fill(deque<int>)/1048576       379841 ns         80079 ns        4.7x
std::fill(vector<int>)/32             1.99 ns          1.79 ns        1.1x
std::fill(vector<int>)/50             3.05 ns          3.06 ns        1.0x
std::fill(vector<int>)/1024           37.6 ns          38.0 ns        1.0x
std::fill(vector<int>)/8192            255 ns           257 ns        1.0x
std::fill(vector<int>)/65536          2966 ns          2981 ns        1.0x
std::fill(vector<int>)/1048576       78300 ns         80348 ns        1.0x
rng::fill(vector<int>)/32             1.77 ns          1.75 ns        1.0x
rng::fill(vector<int>)/50             4.85 ns          2.31 ns        2.1x
rng::fill(vector<int>)/1024           39.6 ns          36.1 ns        1.1x
rng::fill(vector<int>)/8192            238 ns           251 ns        0.9x
rng::fill(vector<int>)/65536          2941 ns          2918 ns        1.0x
rng::fill(vector<int>)/1048576       80497 ns         80442 ns        1.0x
```

---------

Co-authored-by: Louis Dionne <ldionne.2@gmail.com>
Co-authored-by: A. Jiang <de34@live.cn>
2025-10-17 07:41:24 +08:00

118 lines
4.3 KiB
C++

//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
#ifndef _LIBCPP___ALGORITHM_FILL_N_H
#define _LIBCPP___ALGORITHM_FILL_N_H
#include <__algorithm/for_each_n_segment.h>
#include <__algorithm/min.h>
#include <__config>
#include <__fwd/bit_reference.h>
#include <__iterator/iterator_traits.h>
#include <__iterator/segmented_iterator.h>
#include <__memory/pointer_traits.h>
#include <__type_traits/conjunction.h>
#include <__type_traits/enable_if.h>
#include <__type_traits/integral_constant.h>
#include <__type_traits/negation.h>
#include <__utility/convert_to_integral.h>
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
# pragma GCC system_header
#endif
_LIBCPP_PUSH_MACROS
#include <__undef_macros>
_LIBCPP_BEGIN_NAMESPACE_STD
// fill_n isn't specialized for std::memset, because the compiler already optimizes the loop to a call to std::memset.
template <class _OutputIterator,
class _Size,
class _Tp
#ifndef _LIBCPP_CXX03_LANG
,
__enable_if_t<!_And<_BoolConstant<__is_segmented_iterator_v<_OutputIterator>>,
__has_random_access_local_iterator<_OutputIterator>>::value,
int> = 0
#endif
>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _OutputIterator
__fill_n(_OutputIterator __first, _Size __n, const _Tp& __value) {
for (; __n > 0; ++__first, (void)--__n)
*__first = __value;
return __first;
}
#ifndef _LIBCPP_CXX03_LANG
template < class _OutputIterator,
class _Size,
class _Tp,
__enable_if_t<_And<_BoolConstant<__is_segmented_iterator_v<_OutputIterator>>,
__has_random_access_local_iterator<_OutputIterator>>::value,
int> = 0>
inline _LIBCPP_HIDE_FROM_ABI
_LIBCPP_CONSTEXPR_SINCE_CXX14 _OutputIterator __fill_n(_OutputIterator __first, _Size __n, const _Tp& __value) {
using __local_iterator_t = typename __segmented_iterator_traits<_OutputIterator>::__local_iterator;
return std::__for_each_n_segment(__first, __n, [&](__local_iterator_t __lfirst, __local_iterator_t __llast) {
std::__fill_n(__lfirst, __llast - __lfirst, __value);
});
}
#endif // !_LIBCPP_CXX03_LANG
template <bool _FillVal, class _Cp>
_LIBCPP_CONSTEXPR_SINCE_CXX20 _LIBCPP_HIDE_FROM_ABI void
__fill_n_bool(__bit_iterator<_Cp, false> __first, typename __size_difference_type_traits<_Cp>::size_type __n) {
using _It = __bit_iterator<_Cp, false>;
using __storage_type = typename _It::__storage_type;
const int __bits_per_word = _It::__bits_per_word;
// do first partial word
if (__first.__ctz_ != 0) {
__storage_type __clz_f = static_cast<__storage_type>(__bits_per_word - __first.__ctz_);
__storage_type __dn = std::min(__clz_f, __n);
std::__fill_masked_range(std::__to_address(__first.__seg_), __clz_f - __dn, __first.__ctz_, _FillVal);
__n -= __dn;
++__first.__seg_;
}
// do middle whole words
__storage_type __nw = __n / __bits_per_word;
std::__fill_n(std::__to_address(__first.__seg_), __nw, _FillVal ? static_cast<__storage_type>(-1) : 0);
__n -= __nw * __bits_per_word;
// do last partial word
if (__n > 0) {
__first.__seg_ += __nw;
std::__fill_masked_range(std::__to_address(__first.__seg_), __bits_per_word - __n, 0u, _FillVal);
}
}
template <class _Cp, class _Size>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 __bit_iterator<_Cp, false>
__fill_n(__bit_iterator<_Cp, false> __first, _Size __n, const bool& __value) {
if (__n > 0) {
if (__value)
std::__fill_n_bool<true>(__first, __n);
else
std::__fill_n_bool<false>(__first, __n);
}
return __first + __n;
}
template <class _OutputIterator, class _Size, class _Tp>
inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _OutputIterator
fill_n(_OutputIterator __first, _Size __n, const _Tp& __value) {
return std::__fill_n(__first, std::__convert_to_integral(__n), __value);
}
_LIBCPP_END_NAMESPACE_STD
_LIBCPP_POP_MACROS
#endif // _LIBCPP___ALGORITHM_FILL_N_H