[libc++] Vectorize mismatch

libcxx/benchmarks/CMakeLists.txt

@@ -183,6 +183,7 @@ set(BENCHMARK_TESTS
     algorithms/make_heap_then_sort_heap.bench.cpp
     algorithms/min.bench.cpp
     algorithms/min_max_element.bench.cpp
+    algorithms/mismatch.bench.cpp
     algorithms/pop_heap.bench.cpp
     algorithms/pstl.stable_sort.bench.cpp
     algorithms/push_heap.bench.cpp

libcxx/benchmarks/algorithms/mismatch.bench.cpp

@@ -0,0 +1,31 @@
+//===----------------------------------------------------------------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#include <algorithm>
+#include <benchmark/benchmark.h>
+#include <random>
+
+// TODO: Look into benchmarking aligned and unaligned memory explicitly
+// (currently things happen to be aligned because they are malloced that way)
+template <class T>
+static void bm_mismatch(benchmark::State& state) {
+  std::vector<T> vec1(state.range(), '1');
+  std::vector<T> vec2(state.range(), '1');
+  std::mt19937_64 rng(std::random_device{}());
+
+  vec1.back() = '2';
+  for (auto _ : state) {
+    benchmark::DoNotOptimize(vec1);
+    benchmark::DoNotOptimize(std::mismatch(vec1.begin(), vec1.end(), vec2.begin()));
+  }
+}
+BENCHMARK(bm_mismatch<char>)->DenseRange(1, 8)->Range(16, 1 << 20);
+BENCHMARK(bm_mismatch<short>)->DenseRange(1, 8)->Range(16, 1 << 20);
+BENCHMARK(bm_mismatch<int>)->DenseRange(1, 8)->Range(16, 1 << 20);
+
+BENCHMARK_MAIN();

libcxx/docs/ReleaseNotes/19.rst

@@ -51,6 +51,8 @@ Improvements and New Features
 - The performance of growing ``std::vector`` has been improved for trivially relocatable types.
 - The performance of ``ranges::fill`` and ``ranges::fill_n`` has been improved for ``vector<bool>::iterator``\s,
   resulting in a performance increase of up to 1400x.
+- The ``std::mismatch`` algorithm has been optimized for integral types, which can lead up to 40x performance
+  improvements.
 
 Deprecations and Removals
 -------------------------

libcxx/include/CMakeLists.txt

@@ -217,6 +217,7 @@ set(files
   __algorithm/shift_right.h
   __algorithm/shuffle.h
   __algorithm/sift_down.h
+  __algorithm/simd_utils.h
   __algorithm/sort.h
   __algorithm/sort_heap.h
   __algorithm/stable_partition.h

libcxx/include/__algorithm/mismatch.h

@@ -11,23 +11,93 @@
 #define _LIBCPP___ALGORITHM_MISMATCH_H
 
 #include <__algorithm/comp.h>
+#include <__algorithm/simd_utils.h>
+#include <__algorithm/unwrap_iter.h>
 #include <__config>
-#include <__iterator/iterator_traits.h>
+#include <__functional/identity.h>
+#include <__type_traits/invoke.h>
+#include <__type_traits/is_constant_evaluated.h>
+#include <__type_traits/is_equality_comparable.h>
+#include <__type_traits/operation_traits.h>
+#include <__utility/move.h>
 #include <__utility/pair.h>
+#include <__utility/unreachable.h>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header
 #endif
 
+_LIBCPP_PUSH_MACROS
+#include <__undef_macros>
+
 _LIBCPP_BEGIN_NAMESPACE_STD
 
+template <class _Iter1, class _Sent1, class _Iter2, class _Pred, class _Proj1, class _Proj2>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Iter1, _Iter2>
+__mismatch_loop(_Iter1 __first1, _Sent1 __last1, _Iter2 __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  while (__first1 != __last1) {
+    if (!std::__invoke(__pred, std::__invoke(__proj1, *__first1), std::__invoke(__proj2, *__first2)))
+      break;
+    ++__first1;
+    ++__first2;
+  }
+  return std::make_pair(std::move(__first1), std::move(__first2));
+}
+
+template <class _Iter1, class _Sent1, class _Iter2, class _Pred, class _Proj1, class _Proj2>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Iter1, _Iter2>
+__mismatch(_Iter1 __first1, _Sent1 __last1, _Iter2 __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  return std::__mismatch_loop(__first1, __last1, __first2, __pred, __proj1, __proj2);
+}
+
+#if _LIBCPP_VECTORIZE_ALGORITHMS
+
+template <class _Tp,
+          class _Pred,
+          class _Proj1,
+          class _Proj2,
+          __enable_if_t<is_integral<_Tp>::value && __desugars_to<__equal_tag, _Pred, _Tp, _Tp>::value &&
+                            __is_identity<_Proj1>::value && __is_identity<_Proj2>::value,
+                        int> = 0>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_Tp*, _Tp*>
+__mismatch(_Tp* __first1, _Tp* __last1, _Tp* __first2, _Pred& __pred, _Proj1& __proj1, _Proj2& __proj2) {
+  constexpr size_t __unroll_count = 4;
+  constexpr size_t __vec_size     = __native_vector_size<_Tp>;
+  using __vec                     = __simd_vector<_Tp, __vec_size>;
+  if (!__libcpp_is_constant_evaluated()) {
+    while (static_cast<size_t>(__last1 - __first1) >= __unroll_count * __vec_size) [[__unlikely__]] {
+      __vec __lhs[__unroll_count];
+      __vec __rhs[__unroll_count];
+
+      for (size_t __i = 0; __i != __unroll_count; ++__i) {
+        __lhs[__i] = std::__load_vector<__vec>(__first1 + __i * __vec_size);
+        __rhs[__i] = std::__load_vector<__vec>(__first2 + __i * __vec_size);
+      }
+
+      for (size_t __i = 0; __i != __unroll_count; ++__i) {
+        if (auto __cmp_res = __lhs[__i] == __rhs[__i]; !std::__all_of(__cmp_res)) {
+          auto __offset = __i * __vec_size + std::__find_first_not_set(__cmp_res);
+          return {__first1 + __offset, __first2 + __offset};
+        }
+      }
+
+      __first1 += __unroll_count * __vec_size;
+      __first2 += __unroll_count * __vec_size;
+    }
+  }
+  // TODO: Consider vectorizing the tail
+  return std::__mismatch_loop(__first1, __last1, __first2, __pred, __proj1, __proj2);
+}
+
+#endif // _LIBCPP_VECTORIZE_ALGORITHMS
+
 template <class _InputIterator1, class _InputIterator2, class _BinaryPredicate>
 _LIBCPP_NODISCARD_EXT inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 pair<_InputIterator1, _InputIterator2>
 mismatch(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _BinaryPredicate __pred) {
-  for (; __first1 != __last1; ++__first1, (void)++__first2)
-    if (!__pred(*__first1, *__first2))
-      break;
-  return pair<_InputIterator1, _InputIterator2>(__first1, __first2);
+  __identity __proj;
+  auto __res = std::__mismatch(
+      std::__unwrap_iter(__first1), std::__unwrap_iter(__last1), std::__unwrap_iter(__first2), __pred, __proj, __proj);
+  return std::make_pair(std::__rewrap_iter(__first1, __res.first), std::__rewrap_iter(__first2, __res.second));
 }
 
 template <class _InputIterator1, class _InputIterator2>
@@ -59,4 +129,6 @@ mismatch(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __fi
 
 _LIBCPP_END_NAMESPACE_STD
 
+_LIBCPP_POP_MACROS
+
 #endif // _LIBCPP___ALGORITHM_MISMATCH_H

libcxx/include/__algorithm/simd_utils.h

@@ -0,0 +1,123 @@
+//===----------------------------------------------------------------------===//
+//
+// 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_SIMD_UTILS_H
+#define _LIBCPP___ALGORITHM_SIMD_UTILS_H
+
+#include <__bit/bit_cast.h>
+#include <__bit/countr.h>
+#include <__config>
+#include <__type_traits/is_arithmetic.h>
+#include <__type_traits/is_same.h>
+#include <__utility/integer_sequence.h>
+#include <cstddef>
+#include <cstdint>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+#  pragma GCC system_header
+#endif
+
+// TODO: Find out how altivec changes things and allow vectorizations there too.
+#if _LIBCPP_STD_VER >= 14 && defined(_LIBCPP_CLANG_VER) && _LIBCPP_CLANG_VER >= 1700 && !defined(__ALTIVEC__)
+#  define _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS 1
+#else
+#  define _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS 0
+#endif
+
+#if _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS && !defined(__OPTIMIZE_SIZE__)
+#  define _LIBCPP_VECTORIZE_ALGORITHMS 1
+#else
+#  define _LIBCPP_VECTORIZE_ALGORITHMS 0
+#endif
+
+#if _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+// This isn't specialized for 64 byte vectors on purpose. They have the potential to significantly reduce performance
+// in mixed simd/non-simd workloads and don't provide any performance improvement for currently vectorized algorithms
+// as far as benchmarks are concerned.
+#  if defined(__AVX__)
+template <class _Tp>
+inline constexpr size_t __native_vector_size = 32 / sizeof(_Tp);
+#  elif defined(__SSE__) || defined(__ARM_NEON__)
+template <class _Tp>
+inline constexpr size_t __native_vector_size = 16 / sizeof(_Tp);
+#  elif defined(__MMX__)
+template <class _Tp>
+inline constexpr size_t __native_vector_size = 8 / sizeof(_Tp);
+#  else
+template <class _Tp>
+inline constexpr size_t __native_vector_size = 1;
+#  endif
+
+template <class _ArithmeticT, size_t _Np>
+using __simd_vector __attribute__((__ext_vector_type__(_Np))) = _ArithmeticT;
+
+template <class _VecT>
+inline constexpr size_t __simd_vector_size_v = []<bool _False = false>() -> size_t {
+  static_assert(_False, "Not a vector!");
+}();
+
+template <class _Tp, size_t _Np>
+inline constexpr size_t __simd_vector_size_v<__simd_vector<_Tp, _Np>> = _Np;
+
+template <class _Tp, size_t _Np>
+_LIBCPP_HIDE_FROM_ABI _Tp __simd_vector_underlying_type_impl(__simd_vector<_Tp, _Np>) {
+  return _Tp{};
+}
+
+template <class _VecT>
+using __simd_vector_underlying_type_t = decltype(std::__simd_vector_underlying_type_impl(_VecT{}));
+
+// This isn't inlined without always_inline when loading chars.
+template <class _VecT, class _Tp>
+_LIBCPP_NODISCARD _LIBCPP_ALWAYS_INLINE _LIBCPP_HIDE_FROM_ABI _VecT __load_vector(const _Tp* __ptr) noexcept {
+  return [=]<size_t... _Indices>(index_sequence<_Indices...>) _LIBCPP_ALWAYS_INLINE noexcept {
+    return _VecT{__ptr[_Indices]...};
+  }(make_index_sequence<__simd_vector_size_v<_VecT>>{});
+}
+
+template <class _Tp, size_t _Np>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI bool __all_of(__simd_vector<_Tp, _Np> __vec) noexcept {
+  return __builtin_reduce_and(__builtin_convertvector(__vec, __simd_vector<bool, _Np>));
+}
+
+template <class _Tp, size_t _Np>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI size_t __find_first_set(__simd_vector<_Tp, _Np> __vec) noexcept {
+  using __mask_vec = __simd_vector<bool, _Np>;
+
+  // This has MSan disabled du to https://github.com/llvm/llvm-project/issues/85876
+  auto __impl = [&]<class _MaskT>(_MaskT) _LIBCPP_NO_SANITIZE("memory") noexcept {
+    return std::__countr_zero(__builtin_bit_cast(_MaskT, __builtin_convertvector(__vec, __mask_vec)));
+  };
+
+  if constexpr (sizeof(__mask_vec) == sizeof(uint8_t)) {
+    return __impl(uint8_t{});
+  } else if constexpr (sizeof(__mask_vec) == sizeof(uint16_t)) {
+    return __impl(uint16_t{});
+  } else if constexpr (sizeof(__mask_vec) == sizeof(uint32_t)) {
+    return __impl(uint32_t{});
+  } else if constexpr (sizeof(__mask_vec) == sizeof(uint64_t)) {
+    return __impl(uint64_t{});
+  } else {
+    static_assert(sizeof(__mask_vec) == 0, "unexpected required size for mask integer type");
+    return 0;
+  }
+}
+
+template <class _Tp, size_t _Np>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI size_t __find_first_not_set(__simd_vector<_Tp, _Np> __vec) noexcept {
+  return std::__find_first_set(~__vec);
+}
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // _LIBCPP_HAS_ALGORITHM_VECTOR_UTILS
+
+#endif // _LIBCPP___ALGORITHM_SIMD_UTILS_H

libcxx/include/__bit/bit_cast.h

@@ -19,6 +19,15 @@
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 
+#ifndef _LIBCPP_CXX03_LANG
+
+template <class _ToType, class _FromType>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI constexpr _ToType __bit_cast(const _FromType& __from) noexcept {
+  return __builtin_bit_cast(_ToType, __from);
+}
+
+#endif // _LIBCPP_CXX03_LANG
+
 #if _LIBCPP_STD_VER >= 20
 
 template <class _ToType, class _FromType>

libcxx/include/__bit/countr.h

@@ -35,10 +35,8 @@ _LIBCPP_NODISCARD inline _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR int __libcpp_ct
   return __builtin_ctzll(__x);
 }
 
-#if _LIBCPP_STD_VER >= 20
-
-template <__libcpp_unsigned_integer _Tp>
-_LIBCPP_NODISCARD_EXT _LIBCPP_HIDE_FROM_ABI constexpr int countr_zero(_Tp __t) noexcept {
+template <class _Tp>
+_LIBCPP_NODISCARD _LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 int __countr_zero(_Tp __t) _NOEXCEPT {
   if (__t == 0)
     return numeric_limits<_Tp>::digits;
 
@@ -59,6 +57,13 @@ _LIBCPP_NODISCARD_EXT _LIBCPP_HIDE_FROM_ABI constexpr int countr_zero(_Tp __t) n
   }
 }
 
+#if _LIBCPP_STD_VER >= 20
+
+template <__libcpp_unsigned_integer _Tp>
+_LIBCPP_NODISCARD_EXT _LIBCPP_HIDE_FROM_ABI constexpr int countr_zero(_Tp __t) noexcept {
+  return std::__countr_zero(__t);
+}
+
 template <__libcpp_unsigned_integer _Tp>
 _LIBCPP_NODISCARD_EXT _LIBCPP_HIDE_FROM_ABI constexpr int countr_one(_Tp __t) noexcept {
   return __t != numeric_limits<_Tp>::max() ? std::countr_zero(static_cast<_Tp>(~__t)) : numeric_limits<_Tp>::digits;

libcxx/include/libcxx.imp

@@ -217,6 +217,7 @@
   { include: [ "<__algorithm/shift_right.h>", "private", "<algorithm>", "public" ] },
   { include: [ "<__algorithm/shuffle.h>", "private", "<algorithm>", "public" ] },
   { include: [ "<__algorithm/sift_down.h>", "private", "<algorithm>", "public" ] },
+  { include: [ "<__algorithm/simd_utils.h>", "private", "<algorithm>", "public" ] },
   { include: [ "<__algorithm/sort.h>", "private", "<algorithm>", "public" ] },
   { include: [ "<__algorithm/sort_heap.h>", "private", "<algorithm>", "public" ] },
   { include: [ "<__algorithm/stable_partition.h>", "private", "<algorithm>", "public" ] },

libcxx/include/module.modulemap

@@ -697,7 +697,10 @@ module std_private_algorithm_minmax                                      [system
   export *
 }
 module std_private_algorithm_minmax_element                              [system] { header "__algorithm/minmax_element.h" }
-module std_private_algorithm_mismatch                                    [system] { header "__algorithm/mismatch.h" }
+module std_private_algorithm_mismatch                                    [system] {
+  header "__algorithm/mismatch.h"
+  export std_private_algorithm_simd_utils
+}
 module std_private_algorithm_move                                        [system] { header "__algorithm/move.h" }
 module std_private_algorithm_move_backward                               [system] { header "__algorithm/move_backward.h" }
 module std_private_algorithm_next_permutation                            [system] { header "__algorithm/next_permutation.h" }
@@ -1048,6 +1051,7 @@ module std_private_algorithm_sort                                        [system
   header "__algorithm/sort.h"
   export std_private_debug_utils_strict_weak_ordering_check
 }
+module std_private_algorithm_simd_utils                                  [system] { header "__algorithm/simd_utils.h" }
 module std_private_algorithm_sort_heap                                   [system] { header "__algorithm/sort_heap.h" }
 module std_private_algorithm_stable_partition                            [system] { header "__algorithm/stable_partition.h" }
 module std_private_algorithm_stable_sort                                 [system] { header "__algorithm/stable_sort.h" }