Improve search/find_end perf by dropping memcmp (#4654)

AlexGuteniev · StephanTLavavej · web-flow · commit e3ed20614e89 · 2024-06-17T21:51:00.000-07:00
Co-authored-by: Stephan T. Lavavej &lt;stl@nuwen.net&gt;
diff --git a/benchmarks/CMakeLists.txt b/benchmarks/CMakeLists.txt
@@ -119,6 +119,7 @@ add_benchmark(path_lexically_normal src/path_lexically_normal.cpp)
 add_benchmark(priority_queue_push_range src/priority_queue_push_range.cpp)
 add_benchmark(random_integer_generation src/random_integer_generation.cpp)
 add_benchmark(replace src/replace.cpp)
+add_benchmark(search src/search.cpp)
 add_benchmark(std_copy src/std_copy.cpp)
 add_benchmark(swap_ranges src/swap_ranges.cpp)
 
diff --git a/benchmarks/src/replace.cpp b/benchmarks/src/replace.cpp
@@ -10,7 +10,7 @@ const char src[] =
     "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam mollis imperdiet massa, at dapibus elit interdum "
     "ac. In eget sollicitudin mi. Nam at tellus at sapien tincidunt sollicitudin vel non eros. Pellentesque nunc nunc, "
     "ullamcorper eu accumsan at, pulvinar non turpis. Quisque vel mauris pulvinar, pretium purus vel, ultricies erat. "
-    "Curabitur a magna in ligula tristique ornare. Quisque commodo, massa viverra laoreet luctus, sem nisi aliquet "
+    "Curabitur a magna in ligula tristique ornare. Quisque commodo, massa viverra laoreet luctus, sem nisi aliquam "
     "velit, fermentum pulvinar velit leo eget justo. Suspendisse vel erat efficitur, pulvinar eros volutpat, vulputate "
     "ex. Phasellus non purus vel velit tristique tristique id at ligula. Quisque mollis sodales magna. Mauris et quam "
     "eu quam viverra tempus. Nullam tempus maximus porta. Nunc mattis eleifend fermentum. Nullam aliquam libero "
@@ -34,7 +34,7 @@ const char src[] =
     "montes, nascetur ridiculus mus. Mauris varius dui sit amet tortor facilisis vestibulum. Curabitur condimentum "
     "justo nec orci mattis auctor. Quisque aliquet condimentum arcu ac sollicitudin. Maecenas elit elit, condimentum "
     "vitae auctor a, cursus et sem. Cras vehicula ante in consequat fermentum. Praesent at massa nisi. Mauris pretium "
-    "euismod eros, ut posuere ligula ullamcorper id. Nullam aliquam malesuada est at dignissim. Pellentesque finibus "
+    "euismod eros, ut posuere ligula ullamcorper id. Nullam aliquet malesuada est at dignissim. Pellentesque finibus "
     "sagittis libero nec bibendum. Phasellus dolor ipsum, finibus quis turpis quis, mollis interdum felis.";
 
 template <class T>
diff --git a/benchmarks/src/search.cpp b/benchmarks/src/search.cpp
@@ -0,0 +1,114 @@
+// Copyright (c) Microsoft Corporation.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#include <algorithm>
+#include <benchmark/benchmark.h>
+#include <cstdint>
+#include <cstring>
+#include <functional>
+#include <string>
+#include <vector>
+
+const char src_haystack[] =
+    "Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam mollis imperdiet massa, at dapibus elit interdum "
+    "ac. In eget sollicitudin mi. Nam at tellus at sapien tincidunt sollicitudin vel non eros. Pellentesque nunc nunc, "
+    "ullamcorper eu accumsan at, pulvinar non turpis. Quisque vel mauris pulvinar, pretium purus vel, ultricies erat. "
+    "Curabitur a magna in ligula tristique ornare. Quisque commodo, massa viverra laoreet luctus, sem nisi aliquam "
+    "velit, fermentum pulvinar velit leo eget justo. Suspendisse vel erat efficitur, pulvinar eros volutpat, vulputate "
+    "ex. Phasellus non purus vel velit tristique tristique id at ligula. Quisque mollis sodales magna. Mauris et quam "
+    "eu quam viverra tempus. Nullam tempus maximus porta. Nunc mattis eleifend fermentum. Nullam aliquam libero "
+    "accumsan velit elementum, eu laoreet metus convallis. Donec pellentesque lacus ut iaculis iaculis. Curabitur orci "
+    "elit, bibendum sit amet feugiat at, iaculis sit amet massa. Maecenas imperdiet lacus at vehicula iaculis. Donec "
+    "volutpat nunc sit amet accumsan tempor. Quisque pretium vestibulum ultricies. Suspendisse potenti. Aenean at diam "
+    "iaculis, condimentum felis venenatis, condimentum erat. Nam quis elit dui. Duis quis odio vitae metus hendrerit "
+    "rhoncus ut et magna. Cras ac augue quis nibh pharetra sagittis. Donec ullamcorper vel eros semper pretium. Proin "
+    "vel sollicitudin eros. Nulla sollicitudin mattis turpis id suscipit. Aliquam sed risus velit. Aliquam iaculis nec "
+    "nibh ac egestas. Duis finibus semper est sed consequat. Sed in sapien quis nibh dignissim mattis. Vestibulum nec "
+    "metus sodales, euismod mauris ac, sollicitudin libero. Maecenas non arcu ac velit ullamcorper fringilla et quis "
+    "nulla. Curabitur posuere leo eget ipsum tincidunt dignissim. Cras ultricies suscipit neque, quis suscipit tortor "
+    "venenatis non. Cras nisl mi, bibendum in vulputate quis, vestibulum ornare enim. Nunc hendrerit placerat dui, "
+    "aliquam mollis sem convallis et. Integer vitae urna diam. Phasellus et imperdiet est. Maecenas auctor facilisis "
+    "nibh non commodo. Suspendisse iaculis quam id bibendum feugiat. Pellentesque felis erat, egestas a libero ac, "
+    "laoreet consectetur elit. Cras ut suscipit ex. Etiam gravida sem quis ex porta, eu lacinia tortor fermentum. "
+    "Nulla consequat odio enim, sed condimentum est sagittis a. Quisque nec commodo tellus. Phasellus elementum "
+    "feugiat dolor et feugiat. Praesent sed mattis tortor. In vitae sodales purus. Morbi accumsan, ligula et interdum "
+    "lacinia, leo risus suscipit urna, non luctus mi justo eu ipsum. Curabitur venenatis pretium orci id porttitor. "
+    "Quisque dapibus nisl sit amet elit lobortis sagittis. Orci varius natoque penatibus et magnis dis parturient "
+    "montes, nascetur ridiculus mus. Mauris varius dui sit amet tortor facilisis vestibulum. Curabitur condimentum "
+    "justo nec orci mattis auctor. Quisque aliquet condimentum arcu ac sollicitudin. Maecenas elit elit, condimentum "
+    "vitae auctor a, cursus et sem. Cras vehicula ante in consequat fermentum. Praesent at massa nisi. Mauris pretium "
+    "euismod eros, ut posuere ligula ullamcorper id. Nullam aliquet malesuada est at dignissim. Pellentesque finibus "
+    "sagittis libero nec bibendum. Phasellus dolor ipsum, finibus quis turpis quis, mollis interdum felis.";
+
+const char src_needle[] = "aliquet";
+
+void c_strstr(benchmark::State& state) {
+    const std::string haystack(std::begin(src_haystack), std::end(src_haystack));
+    const std::string needle(std::begin(src_needle), std::end(src_needle));
+
+    for (auto _ : state) {
+        benchmark::DoNotOptimize(haystack);
+        benchmark::DoNotOptimize(needle);
+        auto res = std::strstr(haystack.c_str(), needle.c_str());
+        benchmark::DoNotOptimize(res);
+    }
+}
+
+template <class T>
+void classic_search(benchmark::State& state) {
+    const std::vector<T> haystack(std::begin(src_haystack), std::end(src_haystack));
+    const std::vector<T> needle(std::begin(src_needle), std::end(src_needle));
+
+    for (auto _ : state) {
+        benchmark::DoNotOptimize(haystack);
+        benchmark::DoNotOptimize(needle);
+        auto res = std::search(haystack.begin(), haystack.end(), needle.begin(), needle.end());
+        benchmark::DoNotOptimize(res);
+    }
+}
+
+template <class T>
+void ranges_search(benchmark::State& state) {
+    const std::vector<T> haystack(std::begin(src_haystack), std::end(src_haystack));
+    const std::vector<T> needle(std::begin(src_needle), std::end(src_needle));
+
+    for (auto _ : state) {
+        benchmark::DoNotOptimize(haystack);
+        benchmark::DoNotOptimize(needle);
+        auto res = std::ranges::search(haystack, needle);
+        benchmark::DoNotOptimize(res);
+    }
+}
+
+template <class T>
+void search_default_searcher(benchmark::State& state) {
+    const std::vector<T> haystack(std::begin(src_haystack), std::end(src_haystack));
+    const std::vector<T> needle(std::begin(src_needle), std::end(src_needle));
+
+    for (auto _ : state) {
+        benchmark::DoNotOptimize(haystack);
+        benchmark::DoNotOptimize(needle);
+        auto res = std::search(haystack.begin(), haystack.end(), std::default_searcher{needle.begin(), needle.end()});
+        benchmark::DoNotOptimize(res);
+    }
+}
+
+BENCHMARK(c_strstr);
+
+BENCHMARK(classic_search<std::uint8_t>);
+BENCHMARK(classic_search<std::uint16_t>);
+BENCHMARK(classic_search<std::uint32_t>);
+BENCHMARK(classic_search<std::uint64_t>);
+
+BENCHMARK(ranges_search<std::uint8_t>);
+BENCHMARK(ranges_search<std::uint16_t>);
+BENCHMARK(ranges_search<std::uint32_t>);
+BENCHMARK(ranges_search<std::uint64_t>);
+
+BENCHMARK(search_default_searcher<std::uint8_t>);
+BENCHMARK(search_default_searcher<std::uint16_t>);
+BENCHMARK(search_default_searcher<std::uint32_t>);
+BENCHMARK(search_default_searcher<std::uint64_t>);
+
+
+BENCHMARK_MAIN();
diff --git a/stl/inc/algorithm b/stl/inc/algorithm
@@ -200,6 +200,7 @@ _Ty* _Find_last_vectorized(_Ty* const _First, _Ty* const _Last, const _TVal _Val
 template <class _Ty1, class _Ty2>
 _Ty1* _Find_first_of_vectorized(
     _Ty1* const _First1, _Ty1* const _Last1, _Ty2* const _First2, _Ty2* const _Last2) noexcept {
+    _STL_INTERNAL_STATIC_ASSERT(sizeof(_Ty1) == sizeof(_Ty2));
     if constexpr (sizeof(_Ty1) == 1) {
         return const_cast<_Ty1*>(
             static_cast<const _Ty1*>(::__std_find_first_of_trivial_1(_First1, _Last1, _First2, _Last2)));
@@ -2119,15 +2120,6 @@ namespace ranges {
 template <class _InIt1, class _InIt2, class _Pr>
 _NODISCARD _CONSTEXPR20 bool _Equal_rev_pred_unchecked(_InIt1 _First1, _InIt2 _First2, const _InIt2 _Last2, _Pr _Pred) {
     // compare [_First1, ...) to [_First2, _Last2)
-    if constexpr (_Equal_memcmp_is_safe<_InIt1, _InIt2, _Pr>) {
-#if _HAS_CXX20
-        if (!_STD is_constant_evaluated())
-#endif // _HAS_CXX20
-        {
-            return _STD _Memcmp_ranges(_First2, _Last2, _First1) == 0;
-        }
-    }
-
     for (; _First2 != _Last2; ++_First1, (void) ++_First2) {
         if (!_Pred(*_First1, *_First2)) {
             return false;
diff --git a/stl/inc/xutility b/stl/inc/xutility
@@ -5356,7 +5356,7 @@ constexpr bool _Equal_memcmp_is_safe_helper =
 
 template <class _Iter1, class _Iter2, class _Pr>
 constexpr bool _Equal_memcmp_is_safe =
-    _Equal_memcmp_is_safe_helper<remove_const_t<_Iter1>, remove_const_t<_Iter2>, _Pr>;
+    _Equal_memcmp_is_safe_helper<remove_const_t<_Iter1>, remove_const_t<_Iter2>, remove_const_t<_Pr>>;
 
 template <class _CtgIt1, class _CtgIt2>
 _NODISCARD int _Memcmp_ranges(_CtgIt1 _First1, _CtgIt1 _Last1, _CtgIt2 _First2) {
@@ -6648,35 +6648,12 @@ namespace ranges {
 
     _EXPORT_STD inline constexpr _Adjacent_find_fn adjacent_find;
 
-    template <class _It1, class _It2, class _Se2, class _Pr, class _Pj1, class _Pj2>
-    concept _Equal_rev_pred_can_memcmp = is_same_v<_Pj1, identity> && is_same_v<_Pj2, identity>
-                                      && sized_sentinel_for<_Se2, _It2> && _Equal_memcmp_is_safe<_It1, _It2, _Pr>;
-
     template <forward_iterator _It1, input_iterator _It2, sentinel_for<_It2> _Se2, class _Pr, class _Pj1, class _Pj2>
         requires indirectly_comparable<_It1, _It2, _Pr, _Pj1, _Pj2>
     _NODISCARD constexpr pair<bool, _It1> _Equal_rev_pred(
         _It1 _First1, _It2 _First2, const _Se2 _Last2, _Pr _Pred, _Pj1 _Proj1, _Pj2 _Proj2) {
         // Returns {true, _First1 + (_Last2 - _First2)} if [_First1, ...) equals [_First2, _Last2), and {false, {}}
         // otherwise.
-        constexpr bool _Optimize = _Equal_rev_pred_can_memcmp<_It1, _It2, _Se2, _Pr, _Pj1, _Pj2>;
-        if constexpr (_Optimize) {
-            if (!_STD is_constant_evaluated()) {
-                bool _Ans;
-                if constexpr (same_as<_It2, _Se2>) {
-                    _Ans = _STD _Memcmp_ranges(_First2, _Last2, _First1) == 0;
-                } else {
-                    _Ans = _STD _Memcmp_count(_First1, _First2, static_cast<size_t>(_Last2 - _First2)) == 0;
-                }
-
-                if (_Ans) {
-                    _First1 += (_Last2 - _First2);
-                    return {true, _STD move(_First1)};
-                } else {
-                    return {false, _It1 {}};
-                }
-            }
-        }
-
         for (; _First2 != _Last2; ++_First1, (void) ++_First2) {
             if (!_STD invoke(_Pred, _STD invoke(_Proj1, *_First1), _STD invoke(_Proj2, *_First2))) {
                 return {false, _It1 {}};
diff --git a/tests/std/tests/VSO_0000000_vector_algorithms/test.cpp b/tests/std/tests/VSO_0000000_vector_algorithms/test.cpp
@@ -170,6 +170,19 @@ auto last_known_good_find_first_of(FwdItH h_first, FwdItH h_last, FwdItN n_first
     return h_first;
 }
 
+template <class RanItH, class RanItN>
+auto last_known_good_search(RanItH h_first, RanItH h_last, RanItN n_first, RanItN n_last) {
+    const auto n_len = n_last - n_first;
+
+    for (; h_last - h_first >= n_len; ++h_first) {
+        if (equal(h_first, h_first + n_len, n_first, n_last)) {
+            return h_first;
+        }
+    }
+
+    return h_last;
+}
+
 template <class T>
 void test_case_find(const vector<T>& input, T v) {
     auto expected = last_known_good_find(input.begin(), input.end(), v);
@@ -275,12 +288,13 @@ void test_case_find_first_of(const vector<T>& input_haystack, const vector<T>& i
 
 template <class T>
 void test_find_first_of(mt19937_64& gen) {
-    constexpr size_t needleDataCount = 50;
-    using TD                         = conditional_t<sizeof(T) == 1, int, T>;
+    constexpr size_t haystackDataCount = 200;
+    constexpr size_t needleDataCount   = 35;
+    using TD                           = conditional_t<sizeof(T) == 1, int, T>;
     uniform_int_distribution<TD> dis('a', 'z');
     vector<T> input_haystack;
     vector<T> input_needle;
-    input_haystack.reserve(dataCount);
+    input_haystack.reserve(haystackDataCount);
     input_needle.reserve(needleDataCount);
 
     for (;;) {
@@ -292,7 +306,7 @@ void test_find_first_of(mt19937_64& gen) {
             test_case_find_first_of(input_haystack, input_needle);
         }
 
-        if (input_haystack.size() == dataCount) {
+        if (input_haystack.size() == haystackDataCount) {
             break;
         }
 
@@ -312,6 +326,56 @@ void test_find_first_of_containers() {
 #endif // _HAS_CXX20
 }
 
+template <class T>
+void test_case_search(const vector<T>& input_haystack, const vector<T>& input_needle) {
+    auto expected =
+        last_known_good_search(input_haystack.begin(), input_haystack.end(), input_needle.begin(), input_needle.end());
+    auto actual = search(input_haystack.begin(), input_haystack.end(), input_needle.begin(), input_needle.end());
+    assert(expected == actual);
+#if _HAS_CXX17
+    auto searcher_actual = search(
+        input_haystack.begin(), input_haystack.end(), default_searcher{input_needle.begin(), input_needle.end()});
+    assert(expected == searcher_actual);
+#endif // _HAS_CXX17
+#if _HAS_CXX20
+    auto ranges_actual = ranges::search(input_haystack, input_needle);
+    assert(expected == begin(ranges_actual));
+    if (expected != input_haystack.end()) {
+        assert(expected + static_cast<ptrdiff_t>(input_needle.size()) == end(ranges_actual));
+    } else {
+        assert(expected == end(ranges_actual));
+    }
+#endif // _HAS_CXX20
+}
+
+template <class T>
+void test_search(mt19937_64& gen) {
+    constexpr size_t haystackDataCount = 200;
+    constexpr size_t needleDataCount   = 35;
+    using TD                           = conditional_t<sizeof(T) == 1, int, T>;
+    uniform_int_distribution<TD> dis('0', '9');
+    vector<T> input_haystack;
+    vector<T> input_needle;
+    input_haystack.reserve(haystackDataCount);
+    input_needle.reserve(needleDataCount);
+
+    for (;;) {
+        input_needle.clear();
+
+        test_case_search(input_haystack, input_needle);
+        for (size_t attempts = 0; attempts < needleDataCount; ++attempts) {
+            input_needle.push_back(static_cast<T>(dis(gen)));
+            test_case_search(input_haystack, input_needle);
+        }
+
+        if (input_haystack.size() == haystackDataCount) {
+            break;
+        }
+
+        input_haystack.push_back(static_cast<T>(dis(gen)));
+    }
+}
+
 template <class T>
 void test_min_max_element(mt19937_64& gen) {
     using Limits = numeric_limits<T>;
@@ -817,6 +881,16 @@ void test_vector_algorithms(mt19937_64& gen) {
     test_find_first_of_containers<const vector<wchar_t>, vector<wchar_t>>();
     test_find_first_of_containers<vector<char>, vector<int>>();
 
+    test_search<char>(gen);
+    test_search<signed char>(gen);
+    test_search<unsigned char>(gen);
+    test_search<short>(gen);
+    test_search<unsigned short>(gen);
+    test_search<int>(gen);
+    test_search<unsigned int>(gen);
+    test_search<long long>(gen);
+    test_search<unsigned long long>(gen);
+
     test_min_max_element<char>(gen);
     test_min_max_element<signed char>(gen);
     test_min_max_element<unsigned char>(gen);
