无符号数据类型的 MSVC ifstream 性能问题

Question

我在读取二进制文件时用 std::ifstream 在 MSVC 上做了一些测试。我在 char 和 unsigned char 数据类型之间有很大的性能差异。

读取 512 MB 二进制文件时的结果：

Duration read as signed: 322 ms
Duration read as unsigned: 10552 ms

下面是我用来测试的代码：

#include <vector>
#include <iostream>
#include <fstream>
#include <chrono>
#include <limits>
#include <filesystem>

int main()
{
    const std::filesystem::path filePath{ "test.data" }; // 512 MB binary file
    const size_t fileSize{ std::filesystem::file_size(filePath) };

    {
        std::basic_ifstream<char> fileStream{ filePath, std::fstream::binary };
        std::vector<char> data;
        data.resize(fileSize);

        const auto start{ std::chrono::system_clock::now() };
        fileStream.read(data.data(), fileSize);
        const auto end{ std::chrono::system_clock::now() };

        std::cout << "Duration read as signed: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << " ms" << std::endl;
    }

    {
        std::basic_ifstream<unsigned char> fileStream{ filePath, std::fstream::binary };
        std::vector<unsigned char> data;
        data.resize(fileSize);

        const auto start{ std::chrono::system_clock::now() };
        fileStream.read(data.data(), fileSize);
        const auto end{ std::chrono::system_clock::now() };

        std::cout << "Duration read as unsigned: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << " ms" << std::endl;
    }

    return 0;
}

我不明白为什么在读取二进制文件时使用 basic_ifstream<unsigned char> 比 basic_ifstream<char> 慢 30 倍。

Answer 1

我已经追踪到 C:\Program Files (x86)\Microsoft Visual Studio19\Community\VC\Tools\MSVC.29.30037\include\fstream 文件，第 549 行：

virtual streamsize __CLR_OR_THIS_CALL xsgetn(_Elem* _Ptr, streamsize _Count) override {
    // get _Count characters from stream
    if constexpr (sizeof(_Elem) == 1) {
        if (_Count <= 0) {
            return 0;
        }

        if (_Pcvt) { // if we need a nontrivial codecvt transform, do the default expensive thing
            return _Mysb::xsgetn(_Ptr, _Count);
        }

对于 unsigned char 它进入 default expensive thing

再看远一点，我看到了这个：

virtual streamsize __CLR_OR_THIS_CALL xsgetn(_Elem* _Ptr, streamsize _Count) { // get _Count characters from stream
    const streamsize _Start_count = _Count;

    while (0 < _Count) {
        streamsize _Size = _Gnavail();
        if (0 < _Size) { // copy from read buffer
            if (_Count < _Size) {
                _Size = _Count;
            }

            _Traits::copy(_Ptr, gptr(), static_cast<size_t>(_Size));
            _Ptr += _Size;
            _Count -= _Size;
            gbump(static_cast<int>(_Size));
        } else {
            const int_type _Meta = uflow();
            if (_Traits::eq_int_type(_Traits::eof(), _Meta)) {
                break; // end of file, quit
            }

            // get a single character
            *_Ptr++ = _Traits::to_char_type(_Meta);
            --_Count;
        }
    }

    return _Start_count - _Count;
}

备注一一处理！而且那个函数作用不大：

_NODISCARD static constexpr _Elem to_char_type(const int_type& _Meta) noexcept {
    return static_cast<_Elem>(_Meta);
}

Answer 2

当您像这样设置读取缓冲区时，性能问题消失了：

    {
        std::basic_ifstream<unsigned char> fileStream{ filePath, std::fstream::binary };
        std::vector<unsigned char> data;
        data.resize(fileSize);

        unsigned char buf[8192U];
        fileStream.rdbuf()->pubsetbuf(buf, 8192U);

        const auto start{ std::chrono::system_clock::now() };
        fileStream.read(data.data(), fileSize);
        const auto end{ std::chrono::system_clock::now() };

        std::cout << "Duration read unsigned with buffer: " << std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count() << " ms" << std::endl;
    }

结果：

Duration read signed: 331 ms
Duration read unsigned: 10505 ms
Duration read unsigned with buffer: 223 ms

无符号数据类型的 MSVC ifstream 性能问题

MSVC ifstream performance issue with unsigned datatype

c++

visual-c++

ifstream