引入 WIL

[Blinue]

WIL 包含了实用的包装器来管理 Win32 资源、调用函数等，可以大幅简化我们的代码。很多自定义类在 WIL 有替代品。

现有类/函数	WIL 替代品
ScopeExit	RAII 包装器
ScopedHandle	RAII 包装器
CSMutex	critical_section
SRWMutex	srwlocks
BStr	unique_bstr
CreateDir	CreateDirectoryDeepNoThrow

其他更改：

1. 使用惰性求值优化配置文件的匹配性能
2. 优化大量错误处理
3. 新增 WinRTUtils::Event 简化事件创建
4. 优化下载更新的性能

[Blinue]

我用下面的代码测试了 CNG 和 Windows.Security.Cryptography (以下简称 WSC) 的哈希性能。发现 CNG 比 WSC 快了一个数量级。

#define NOMINMAX
#define WIN32_NO_STATUS

#include <iostream>
#include <Windows.h>
#include <winrt/Windows.Security.Cryptography.Core.h>
#include <winrt/Windows.Storage.Streams.h>
#include <bcrypt.h>
#include <ntstatus.h>

template<typename Fn>
static int Measure(const Fn& func) {
    using namespace std::chrono;

    auto t = steady_clock::now();
    func();
    auto duration = steady_clock::now() - t;

    return int(duration_cast<microseconds>(duration).count());
}

static void Win32Hash(const std::vector<uint8_t>& data, std::array<uint8_t, 16>& hash) {
    BCRYPT_ALG_HANDLE hAlg = NULL;
    BCryptOpenAlgorithmProvider(&hAlg, BCRYPT_MD5_ALGORITHM, nullptr, 0);

    uint32_t hashObjSize = 0;
    ULONG result;
    BCryptGetProperty(hAlg, BCRYPT_OBJECT_LENGTH, (PUCHAR)&hashObjSize, sizeof(hashObjSize), &result, 0);

    std::unique_ptr<uint8_t[]> hashObj = std::make_unique<uint8_t[]>(hashObjSize);

    BCRYPT_HASH_HANDLE hHash = NULL;
    BCryptCreateHash(hAlg, &hHash, hashObj.get(), hashObjSize, NULL, 0, 0);

    for (size_t i = 0; i < data.size(); i += 100) {
        BCryptHashData(hHash, (PUCHAR)data.data() + i, std::min(ULONG(100), ULONG(data.size() - i)), 0);
    }

    BCryptFinishHash(hHash, hash.data(), (ULONG)hash.size(), 0);

    BCryptDestroyHash(hHash);
    BCryptCloseAlgorithmProvider(hAlg, 0);
}

struct __declspec(uuid("905a0fef-bc53-11df-8c49-001e4fc686da")) IBufferByteAccess : IUnknown {
    virtual HRESULT __stdcall Buffer(void** value) = 0;
};

struct CustomBuffer : winrt::implements<CustomBuffer, winrt::Windows::Storage::Streams::IBuffer, IBufferByteAccess> {
    void* _address;
    uint32_t _size;

    CustomBuffer(void* address, uint32_t size) : _address(address), _size(size) {}
    uint32_t Capacity() const { return _size; }
    uint32_t Length() const { return _size; }
    void Length(uint32_t length) { throw winrt::hresult_not_implemented(); }

    HRESULT __stdcall Buffer(void** value) final {
        *value = _address;
        return S_OK;
    }
};

static void WinRTHash(const std::vector<uint8_t>& data, std::array<uint8_t, 16>& hash) {
    using namespace winrt;
    using namespace Windows::Security::Cryptography::Core;
    using namespace Windows::Storage::Streams;

    HashAlgorithmProvider hashAlgProvider = HashAlgorithmProvider::OpenAlgorithm(HashAlgorithmNames::Md5());
    CryptographicHash hasher = hashAlgProvider.CreateHash();

    for (size_t i = 0; i < data.size(); i += 100) {
        IBuffer buffer = make<CustomBuffer>((void*)(data.data() + i), std::min(100u, uint32_t(data.size() - i)));
        hasher.Append(buffer);
    }

    IBuffer hashBuffer = hasher.GetValueAndReset();
    assert(hashBuffer.Length() == hash.size());
    std::memcpy(hash.data(), hashBuffer.data(), hash.size());
}

static std::string HexHash(std::array<uint8_t, 16> hash) {
    static char oct2Hex[16] = {
        '0','1','2','3','4','5','6','7',
        '8','9','a','b','c','d','e','f'
    };

    std::string result(16, 0);
    char* pResult = &result[0];

    uint8_t* b = hash.data();
    for (int i = 0; i < hash.size(); ++i) {
        *pResult++ = oct2Hex[(*b >> 4) & 0xf];
        *pResult++ = oct2Hex[*b & 0xf];
        ++b;
    }

    return result;
}

int main() {
    std::vector<uint8_t> data(100000);
    std::generate(data.begin(), data.end(), []() { return (uint8_t)std::rand(); });

    std::array<uint8_t, 16> hash;

    int t = Measure([&]() {
        Win32Hash(data, hash);
    });
    std::cout << "Win32: " << HexHash(hash) << " " << t << "us\n";

    t = Measure([&]() {
        WinRTHash(data, hash);
    });
    std::cout << "WinRT: " << HexHash(hash) << " " << t << "us" << std::endl;
}

代码中 CustomBuffer 用于将内存块传给 WinRT 而避免额外复制。

测试结果如下：

数据量 (KiB)	CNG (us)	WSC (us)
1	229	2736
10	239	2671
100	488	2978
1000	1983	5531

WSC 如此慢的主要原因是它需要加载数个 dll，包括 kernel.appcore.dll、bcryptprimitives.dll、clbcatq.dll、CryptoWinRT.dll、WinTypes.dll，而 CNG 只需加载 bcryptprimitives.dll。显然 WSC 只是 CNG 的很慢的包装。

接下来先预热一次来比较真正的执行性能。

更新 Measure 函数：

template<typename Fn>
static int Measure(const Fn& func) {
    using namespace std::chrono;

    // 预热
    func();

    nanoseconds total{};
    for (int i = 0; i < 10; ++i) {
        auto t = steady_clock::now();
        func();
        total += steady_clock::now() - t;
    }

    return int(duration_cast<microseconds>(total).count() / 10);
}

测试结果如下：

数据量 (KiB)	CNG (us)	WSC (us)
1	2	4
10	17	28
100	165	253
1000	1640	2530

初始化 CNG 需要约 300us，而 WSC 需要约 3000us。即使只比较执行性能，WSC 仍比 CNG 慢约 50%。

结论：避免使用 WSC。事实上，如果有 Win32 替代，应避免使用任何 WinRT API。