[libc++] Optimize vector growing of trivially relocatable types

Author: philnik777

libcxx/benchmarks/ContainerBenchmarks.h

@@ -80,7 +80,7 @@ void BM_ConstructFromRange(benchmark::State& st, Container, GenInputs gen) {
 }
 
 template <class Container>
-void BM_Pushback(benchmark::State& state, Container c) {
+void BM_Pushback_no_grow(benchmark::State& state, Container c) {
   int count = state.range(0);
   c.reserve(count);
   while (state.KeepRunningBatch(count)) {

libcxx/benchmarks/vector_operations.bench.cpp

@@ -1,6 +1,7 @@
 #include <cstdint>
 #include <cstdlib>
 #include <cstring>
+#include <deque>
 #include <functional>
 #include <vector>
 
@@ -39,6 +40,21 @@ BENCHMARK_CAPTURE(BM_ConstructFromRange, vector_size_t, std::vector<size_t>{}, g
 BENCHMARK_CAPTURE(BM_ConstructFromRange, vector_string, std::vector<std::string>{}, getRandomStringInputs)
     ->Arg(TestNumInputs);
 
-BENCHMARK_CAPTURE(BM_Pushback, vector_int, std::vector<int>{})->Arg(TestNumInputs);
+BENCHMARK_CAPTURE(BM_Pushback_no_grow, vector_int, std::vector<int>{})->Arg(TestNumInputs);
+
+template <class T>
+void bm_grow(benchmark::State& state) {
+  for (auto _ : state) {
+    std::vector<T> vec;
+    benchmark::DoNotOptimize(vec);
+    for (size_t i = 0; i != 2048; ++i)
+      vec.emplace_back();
+    benchmark::DoNotOptimize(vec);
+  }
+}
+BENCHMARK(bm_grow<int>);
+BENCHMARK(bm_grow<std::string>);
+BENCHMARK(bm_grow<std::unique_ptr<int>>);
+BENCHMARK(bm_grow<std::deque<int>>);
 
 BENCHMARK_MAIN();

libcxx/docs/ReleaseNotes/18.rst

@@ -69,6 +69,8 @@ Improvements and New Features
 - ``std::for_each`` has been optimized for segmented iterators like ``std::deque::iterator`` in C++23 and
   later, which can lead up to 40x performance improvements.
 
+- The performance of growing ``std::vector`` has been improved for trivially relocatable types.
+
 - The library now provides several hardening modes under which common cases of library undefined behavior will be turned
   into a reliable program termination. The ``fast`` hardening mode enables a set of security-critical checks with
   minimal runtime overhead; the ``extensive`` hardening mode additionally enables relatively cheap checks that catch

libcxx/include/CMakeLists.txt

@@ -803,6 +803,7 @@ set(files
   __type_traits/is_trivially_lexicographically_comparable.h
   __type_traits/is_trivially_move_assignable.h
   __type_traits/is_trivially_move_constructible.h
+  __type_traits/is_trivially_relocatable.h
   __type_traits/is_unbounded_array.h
   __type_traits/is_union.h
   __type_traits/is_unsigned.h

libcxx/include/__fwd/string.h

@@ -12,6 +12,8 @@
 #include <__availability>
 #include <__config>
 #include <__fwd/memory_resource.h>
+#include <__type_traits/is_fundamental.h>
+#include <__type_traits/is_trivially_relocatable.h>
 
 #if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
 #  pragma GCC system_header

libcxx/include/__memory/uninitialized_algorithms.h

@@ -29,6 +29,7 @@
 #include <__type_traits/is_trivially_copy_constructible.h>
 #include <__type_traits/is_trivially_move_assignable.h>
 #include <__type_traits/is_trivially_move_constructible.h>
+#include <__type_traits/is_trivially_relocatable.h>
 #include <__type_traits/is_unbounded_array.h>
 #include <__type_traits/negation.h>
 #include <__type_traits/remove_const.h>
@@ -591,60 +592,48 @@ __uninitialized_allocator_copy(_Alloc& __alloc, _Iter1 __first1, _Sent1 __last1,
   return std::__rewrap_iter(__first2, __result);
 }
 
-// Move-construct the elements [__first1, __last1) into [__first2, __first2 + N)
-// if the move constructor is noexcept, where N is distance(__first1, __last1).
-//
-// Otherwise try to copy all elements. If an exception is thrown the already copied
-// elements are destroyed in reverse order of their construction.
-template <class _Alloc, class _Iter1, class _Sent1, class _Iter2>
-_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _Iter2
-__uninitialized_allocator_move_if_noexcept(_Alloc& __alloc, _Iter1 __first1, _Sent1 __last1, _Iter2 __first2) {
-  static_assert(__is_cpp17_move_insertable<_Alloc>::value,
-                "The specified type does not meet the requirements of Cpp17MoveInsertable");
-  auto __destruct_first = __first2;
-  auto __guard =
-      std::__make_exception_guard(_AllocatorDestroyRangeReverse<_Alloc, _Iter2>(__alloc, __destruct_first, __first2));
-  while (__first1 != __last1) {
-#ifndef _LIBCPP_HAS_NO_EXCEPTIONS
-    allocator_traits<_Alloc>::construct(__alloc, std::__to_address(__first2), std::move_if_noexcept(*__first1));
-#else
-    allocator_traits<_Alloc>::construct(__alloc, std::__to_address(__first2), std::move(*__first1));
-#endif
-    ++__first1;
-    ++__first2;
-  }
-  __guard.__complete();
-  return __first2;
-}
-
 template <class _Alloc, class _Type>
 struct __allocator_has_trivial_move_construct : _Not<__has_construct<_Alloc, _Type*, _Type&&> > {};
 
 template <class _Type>
 struct __allocator_has_trivial_move_construct<allocator<_Type>, _Type> : true_type {};
 
-#ifndef _LIBCPP_COMPILER_GCC
-template <
-    class _Alloc,
-    class _Iter1,
-    class _Iter2,
-    class _Type = typename iterator_traits<_Iter1>::value_type,
-    class = __enable_if_t<is_trivially_move_constructible<_Type>::value && is_trivially_move_assignable<_Type>::value &&
-                          __allocator_has_trivial_move_construct<_Alloc, _Type>::value> >
-_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX20 _Iter2
-__uninitialized_allocator_move_if_noexcept(_Alloc&, _Iter1 __first1, _Iter1 __last1, _Iter2 __first2) {
-  if (__libcpp_is_constant_evaluated()) {
-    while (__first1 != __last1) {
-      std::__construct_at(std::__to_address(__first2), std::move(*__first1));
-      ++__first1;
-      ++__first2;
+template <class _Alloc, class _Tp>
+struct __allocator_has_trivial_destroy : _Not<__has_destroy<_Alloc, _Tp*>> {};
+
+template <class _Tp, class _Up>
+struct __allocator_has_trivial_destroy<allocator<_Tp>, _Up> : true_type {};
+
+// __uninitialized_allocator_relocate relocates the objects in [__first, __last) into __result.
+// relocation means that the objects in [__first, __last) are placed into __result as-if by move-construct and destroy,
+// except that the move constructor and destructor may never be called if they are known to be equivalent to a memcpy.
+//
+// Preconditions: __result doesn't contain any objects and [__first, __last) contains objects
+// Postconditions: __result contains the object from [__first, __last) and [__first, __last) doesn't contain any objects
+template <class _Alloc, class _Tp>
+_LIBCPP_HIDE_FROM_ABI _LIBCPP_CONSTEXPR_SINCE_CXX14 void
+__uninitialized_allocator_relocate(_Alloc& __alloc, _Tp* __first, _Tp* __last, _Tp* __result) {
+  if (__libcpp_is_constant_evaluated() || !__libcpp_is_trivially_relocatable<_Tp>::value ||
+      !__allocator_has_trivial_move_construct<_Alloc, _Tp>::value ||
+      !__allocator_has_trivial_destroy<_Alloc, _Tp>::value) {
+    auto __destruct_first = __result;
+    auto __guard =
+        std::__make_exception_guard(_AllocatorDestroyRangeReverse<_Alloc, _Tp*>(__alloc, __destruct_first, __result));
+    while (__first != __last) {
+#ifndef _LIBCPP_HAS_NO_EXCEPTIONS
+      allocator_traits<_Alloc>::construct(__alloc, std::__to_address(__result), std::move_if_noexcept(*__first));
+#else
+      allocator_traits<_Alloc>::construct(__alloc, std::__to_address(__result), std::move(*__first));
+#endif
+      ++__first;
+      ++__result;
     }
-    return __first2;
+    __guard.__complete();
+    std::__allocator_destroy(__alloc, __first, __last);
   } else {
-    return std::move(__first1, __last1, __first2);
+    __builtin_memcpy(__result, __first, sizeof(_Tp) * (__last - __first));
   }
 }
-#endif // _LIBCPP_COMPILER_GCC
 
 _LIBCPP_END_NAMESPACE_STD
 

libcxx/include/__memory/unique_ptr.h

@@ -33,6 +33,7 @@
 #include <__type_traits/is_reference.h>
 #include <__type_traits/is_same.h>
 #include <__type_traits/is_swappable.h>
+#include <__type_traits/is_trivially_relocatable.h>
 #include <__type_traits/is_void.h>
 #include <__type_traits/remove_extent.h>
 #include <__type_traits/type_identity.h>
@@ -129,6 +130,17 @@ class _LIBCPP_UNIQUE_PTR_TRIVIAL_ABI _LIBCPP_TEMPLATE_VIS unique_ptr {
 
   static_assert(!is_rvalue_reference<deleter_type>::value, "the specified deleter type cannot be an rvalue reference");
 
+  // A unique_ptr contains the following members which may be trivially relocatable:
+  // - pointer : this may be trivially relocatable, so it's checked
+  // - delter_type: this may be trivially relocatable, so it's checked
+  //
+  // This uniuqe_ptr implementation only contains a pointer to the unique object and a delter, so there are no
+  // references to itself. This means that the entire structure is trivially relocatable if it's members are.
+  using __trivially_relocatable = __conditional_t<
+      __libcpp_is_trivially_relocatable<pointer>::value && __libcpp_is_trivially_relocatable<deleter_type>::value,
+      unique_ptr,
+      void>;
+
 private:
   __compressed_pair<pointer, deleter_type> __ptr_;
 
@@ -280,6 +292,17 @@ class _LIBCPP_UNIQUE_PTR_TRIVIAL_ABI _LIBCPP_TEMPLATE_VIS unique_ptr<_Tp[], _Dp>
   typedef _Dp deleter_type;
   typedef typename __pointer<_Tp, deleter_type>::type pointer;
 
+  // A unique_ptr contains the following members which may be trivially relocatable:
+  // - pointer : this may be trivially relocatable, so it's checked
+  // - delter_type: this may be trivially relocatable, so it's checked
+  //
+  // This uniuqe_ptr implementation only contains a pointer to the unique object and a delter, so there are no
+  // references to itself. This means that the entire structure is trivially relocatable if it's members are.
+  using __trivially_relocatable = __conditional_t<
+      __libcpp_is_trivially_relocatable<pointer>::value && __libcpp_is_trivially_relocatable<deleter_type>::value,
+      unique_ptr,
+      void>;
+
 private:
   __compressed_pair<pointer, deleter_type> __ptr_;
 

libcxx/include/__type_traits/is_trivially_relocatable.h

@@ -0,0 +1,39 @@
+//===----------------------------------------------------------------------===//
+//
+// 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___TYPE_TRAITS_IS_TRIVIALLY_RELOCATABLE_H
+#define _LIBCPP___TYPE_TRAITS_IS_TRIVIALLY_RELOCATABLE_H
+
+#include <__config>
+#include <__type_traits/enable_if.h>
+#include <__type_traits/integral_constant.h>
+#include <__type_traits/is_same.h>
+#include <__type_traits/is_trivially_copyable.h>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+#  pragma GCC system_header
+#endif
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+#if __has_builtin(__is_trivially_relocatable)
+template <class _Tp, class = void>
+struct __libcpp_is_trivially_relocatable : integral_constant<bool, __is_trivially_relocatable(_Tp)> {};
+#else
+template <class _Tp, class = void>
+struct __libcpp_is_trivially_relocatable : is_trivially_copyable<_Tp> {};
+#endif
+
+template <class _Tp>
+struct __libcpp_is_trivially_relocatable<_Tp,
+                                         __enable_if_t<is_same<_Tp, typename _Tp::__trivially_relocatable>::value> >
+    : true_type {};
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // _LIBCPP___TYPE_TRAITS_IS_TRIVIALLY_RELOCATABLE_H

libcxx/include/string

@@ -724,6 +724,20 @@ public:
   typedef typename __alloc_traits::pointer pointer;
   typedef typename __alloc_traits::const_pointer const_pointer;
 
+  // A basic_string contains the following members which may be trivially relocatable:
+  // - pointer: is currently assumed to be trivially relocatable, but is still checked in case that changes
+  // - size_type: is always trivially relocatable, since it has to be an integral type
+  // - value_type: is always trivially relocatable, since it has to be trivial
+  // - unsigned char: is a fundamental type, so it's trivially relocatable
+  // - allocator_type: may or may not be trivially relocatable, so it's checked
+  //
+  // This string implementation doesn't contain any references into itself. It only contains a bit that says whether
+  // it is in small or large string mode, so the entire structure is trivially relocatable if it's members are.
+  using __trivially_relocatable = __conditional_t<
+      __libcpp_is_trivially_relocatable<allocator_type>::value && __libcpp_is_trivially_relocatable<pointer>::value,
+      basic_string,
+      void>;
+
   static_assert((!is_array<value_type>::value), "Character type of basic_string must not be an array");
   static_assert((is_standard_layout<value_type>::value), "Character type of basic_string must be standard-layout");
   static_assert((is_trivial<value_type>::value), "Character type of basic_string must be trivial");

libcxx/include/vector

@@ -982,37 +982,54 @@ template <ranges::input_range _Range,
 vector(from_range_t, _Range&&, _Alloc = _Alloc()) -> vector<ranges::range_value_t<_Range>, _Alloc>;
 #endif
 
+// __swap_out_circular_buffer relocates the objects in [__begin_, __end_) into the front of __v and swaps the buffers of
+// *this and __v. It is assumed that __v provides space for exactly (__end_ - __begin_) objects in the front. This
+// function has a strong exception guarantee.
 template <class _Tp, class _Allocator>
 _LIBCPP_CONSTEXPR_SINCE_CXX20 void
 vector<_Tp, _Allocator>::__swap_out_circular_buffer(__split_buffer<value_type, allocator_type&>& __v) {
   __annotate_delete();
-  using _RevIter = std::reverse_iterator<pointer>;
-  __v.__begin_   = std::__uninitialized_allocator_move_if_noexcept(
-                     __alloc(), _RevIter(__end_), _RevIter(__begin_), _RevIter(__v.__begin_))
-                     .base();
+  auto __new_begin = __v.__begin_ - (__end_ - __begin_);
+  std::__uninitialized_allocator_relocate(
+      __alloc(), std::__to_address(__begin_), std::__to_address(__end_), std::__to_address(__new_begin));
+  __v.__begin_ = __new_begin;
+  __end_       = __begin_; // All the objects have been destroyed by relocating them.
   std::swap(this->__begin_, __v.__begin_);
   std::swap(this->__end_, __v.__end_);
   std::swap(this->__end_cap(), __v.__end_cap());
   __v.__first_ = __v.__begin_;
   __annotate_new(size());
 }
 
+// __swap_out_circular_buffer relocates the objects in [__begin_, __p) into the front of __v, the objects in
+// [__p, __end_) into the back of __v and swaps the buffers of *this and __v. It is assumed that __v provides space for
+// exactly (__p - __begin_) objects in the front and space for at least (__end_ - __p) objects in the back. This
+// function has a strong exception guarantee if __begin_ == __p || __end_ == __p.
 template <class _Tp, class _Allocator>
 _LIBCPP_CONSTEXPR_SINCE_CXX20 typename vector<_Tp, _Allocator>::pointer
 vector<_Tp, _Allocator>::__swap_out_circular_buffer(__split_buffer<value_type, allocator_type&>& __v, pointer __p) {
   __annotate_delete();
-  pointer __r    = __v.__begin_;
-  using _RevIter = std::reverse_iterator<pointer>;
-  __v.__begin_   = std::__uninitialized_allocator_move_if_noexcept(
-                     __alloc(), _RevIter(__p), _RevIter(__begin_), _RevIter(__v.__begin_))
-                     .base();
-  __v.__end_ = std::__uninitialized_allocator_move_if_noexcept(__alloc(), __p, __end_, __v.__end_);
+  pointer __ret = __v.__begin_;
+
+  // Relocate [__p, __end_) first to avoid having a hole in [__begin_, __end_)
+  // in case something in [__begin_, __p) throws.
+  std::__uninitialized_allocator_relocate(
+      __alloc(), std::__to_address(__p), std::__to_address(__end_), std::__to_address(__v.__end_));
+  __v.__end_ += (__end_ - __p);
+  __end_           = __p; // The objects in [__p, __end_) have been destroyed by relocating them.
+  auto __new_begin = __v.__begin_ - (__p - __begin_);
+
+  std::__uninitialized_allocator_relocate(
+      __alloc(), std::__to_address(__begin_), std::__to_address(__p), std::__to_address(__new_begin));
+  __v.__begin_ = __new_begin;
+  __end_       = __begin_; // All the objects have been destroyed by relocating them.
+
   std::swap(this->__begin_, __v.__begin_);
   std::swap(this->__end_, __v.__end_);
   std::swap(this->__end_cap(), __v.__end_cap());
   __v.__first_ = __v.__begin_;
   __annotate_new(size());
-  return __r;
+  return __ret;
 }
 
 template <class _Tp, class _Allocator>