13  ## Example - memcpy • Problem space • Performance requirements • Scalar, various vectors, specialty instructions, • Alignment: source and destination • Size • Direction • Linearity ## Example – memcpy: scalar naive void *memcpy_scalar(char *dst, const char *src, size_t n) { if(n) { while (n--) { no vectorization 0000000000001270 <_z13memcpy_scalarPcPKcm>: 1270: 48 89 f8 mov rax, rdi 1273: 48 85 d2 test rdx, rdx 1276: 74 1b je 1293 <_z13memcpy_scalarPcPKcm+0x23> 1278: 31 c9 xor ecx

mQ->mWriteCounter.store(mLocalCounter, std::memory_order_release); std::memcpy(mNextElement, &size, sizeof(int32_t)); std::memcpy(mNextElement + sizeof(int32_t), buffer.data(), buffer.size()); mQ->mReadCounter.load(std::memory_order_acquire)) return 0; int32_t size; std::memcpy(&size, mNextElement, sizeof(int32_t)); int32_t writeCounter = mQ->mWriteCounter.load overflow"); EXPECT(size <= buffer.size(), "buffer space isn't large enough"); std::memcpy(buffer.data(), mNextElement + sizeof(size), size); const int32_t payloadSize = sizeof(size)

remote_ptr = ...; // Calculate data auto values = algorithm(1.0f, 3, data); // Send data to proc.1 BCL::memcpy(remote_ptr, values.data(), values.size() * sizeof(float)); BCL::flush(); // Data is copied. ## How t=""> struct GlobalPtr { private: size_t rank_; size_t offset_; }; void memcpy(void* dest, GlobalPtr src, size_t n) { // Issue remote backend::remote_get(dest, src, n, ...); } }; ## Remote Pointer Types - Can build memcpy to support reading/writing from/to remote memory - Can write fetch_and_op, compare_and_swap, etc

remote_ptr = ...; // Calculate data auto values = algorithm(1.0f, 3, data); // Send data to proc.1 memcpy(remote_ptr, values.data(), values.size() * sizeof(float)); flush(); // Data is copied t=""> struct remote_ptr { private: size_t rank_; size_t offset_; }; void memcpy(void* dest, remote_ptr src, size_t count) { // Issue remote get operation to T& operator=(const T& value) { memcpy(ptr_, &value, sizeof(T)); return value; } operator T() { T value; memcpy(&value, ptr_, sizeof(T)); return

are trivial only on MSVC. This is not a problem for the function calls. But a problem for std::memcpy and std::bit_cast. • std::tuple is never trivially move constructible. ## Can we do something about #include void raw_copy(std::byte* dst, std::byte const* src, size_t size) { std::memcpy(dst, src, size); } void checked_copy(// imagine 2 std::spans here std::byte* dst, std::byte const* std::memcpy(dst, src, dst_size); |armv8-a clang 18.1.0|x86-64 gcc 14.2|x64 msvc v19.40 VS17.10|139| |---|---|---|---| |b memcpy|jmp memcpy|jmp memcpy|RAW| |b memcpy|jmp memcpy|jmp

usually a good thing! implicit ones too: e.g. future.get() please, just don't... TracyLockable memcpy(), malloc() et al. shader compilation sorting, searching, container manipulations, ... file, network implicit ones too: e.g. future.get() better yet, avoid sleep() in production code! TracyLockable memcpy(), malloc() et al. shader compilation sorting, searching, container manipulations, ... file, network don't call memcpy() directly... call some my_memcpy() which then calls memcpy() void* my_memcpy(void* dst, const void* src, size_t size) ZoneScopedC(tracy::Color::Blue); return ::memcpy(dst, src, size);

usually a good thing! implicit ones too: e.g. future.get() please, just don't... TracyLockable memcpy(), malloc() et al. shader compilation sorting, searching, container manipulations, ... file, network implicit ones too: e.g. future.get() better yet, avoid sleep() in production code! TracyLockable memcpy(), malloc() et al. shader compilation sorting, searching, container manipulations, ... file, network don't call memcpy() directly... call some my_memcpy() which then calls memcpy() void* my_memcpy(void* dst, const void* src, size_t size) ZoneScopedC(tracy::Color::Blue); return ::memcpy(dst, src, size);

&other) { m_size = other.m_size; m_data = (int *)malloc(m_size * sizeof(int)); memcpy(m_data, other.m_data, m_size * sizeof(int)); }  { m_size = other.m_size; m_data = (int *)malloc(m_size * sizeof(int)); memcpy(m_data, other.m_data, m_size * sizeof(int)); } Vector &operator=(Vector const &other) &other) { m_size = other.m_size; m_data = (int *)realloc(m_data, m_size * sizeof(int)); memcpy(m_data, other.m_data, m_size * sizeof(int)); return *this; } ## C++11：为什么区分拷贝和移动？ - 有时候，我们需要把一个对象

> 0) { // if not empty data = new char[len+1]; // -new memory memcpy(data, s.data, // -copy chars len+1); } } }; }; C++ > 0) { // if not empty data = new char[len+1]; // -new memory memcpy(data, s.data, // -copy chars len+1); } } }; }; public: > 0) { // if not empty data = new char[len+1]; // -new memory memcpy(data, s.data, // -copy chars len+1); } } } };

命名空间中的版本是带有多种重载的。 - 建议别用全局的任何函数（C 语言原始的），始终带上 std:: 前缀（C++ 改良后的）。 • C++ 甚至还有 std::printf, std::memcpy, std::size_t 虽然这些其实没有任何区别..... main.cpp #include #include int main() { float jpg) ## C 语言特色：字符串以 0 结尾 - 这就是为什么 strcpy(dst, src) 只需要两个指针做参数，而 memcpy(dst, src, n) 额外需要一个长度做参数。因为 strcpy 针对的是以 0 结尾的字符数组（C 语言特色），而 memcpy 需要面对的是任意类型的数组。 • strlen(s) 的本质无非是从指针 s 开始，第几个元素是 0，也就是字符串的长度了。