c++ - 循环和if的优化

Question

我有一个程序如下所示:

void Process1(unsigned char* data)
{

}
void Process2(unsigned char* data)
{

}
void Process3(unsigned char* data)
{

}

#define FLAG1 (1 << 1)
#define FLAG2 (1 << 2)
#define FLAG3 (1 << 3)

void ProcessData(unsigned char* data, unsigned int bytes, unsigned int flags)
{
    bool b1 = !!(flags & FLAG1);
    bool b2 = !!(flags & FLAG2);
    bool b3 = !!(flags & FLAG3);
    for (unsigned int i = 0; i < bytes; i ++)
    {
        if (b1) Process1(data + i);
        if (b2) Process2(data + i);
        if (b3) Process3(data + i);
    }
}

看起来，flags & FLAG1 A.K.A b1 在所有循环中都不会改变。但是我们仍然必须在每个循环中进行分支。我只是想知道是否有办法动态地避免这个不必要的分支。

这是 Lundin 解决方案的演示。

#include <windows.h>
#include <stdio.h>
#include <time.h>
LARGE_INTEGER ls, le, ll;
#define START_CLOCK() QueryPerformanceCounter(&ls)
#define END_CLOCK() printf ("%.0lf ns\n", (QueryPerformanceCounter(&le), ((double)le.QuadPart - ls.QuadPart) / ll.QuadPart * 1000000));


void Process1(unsigned char* data)
{
    (*data)++;
}
void Process2(unsigned char* data)
{
    (*data)--;
}
void Process3(unsigned char* data)
{
    (*data) *= (*data);
}

#define FLAG1 (1 << 1)
#define FLAG2 (1 << 2)
#define FLAG3 (1 << 3)

void ProcessData(unsigned char* data, unsigned int bytes, unsigned int flags)
{
    bool b1 = !!(flags & FLAG1);
    bool b2 = !!(flags & FLAG2);
    bool b3 = !!(flags & FLAG3);
    for (unsigned int i = 0; i < bytes; i ++)
    {
        if (b1) Process1(data + i);
        if (b2) Process2(data + i);
        if (b3) Process3(data + i);
    }
}


typedef void (*proc_t)(unsigned char*);

inline static void do_nothing (unsigned char* ptr)
{
    (void)ptr;
}

void ProcessData_x(unsigned char* data, unsigned int bytes, unsigned int flags)
{
    bool b1 = (flags & FLAG1) != 0;  // de-obfuscate the boolean logic
    bool b2 = (flags & FLAG2) != 0;
    bool b3 = (flags & FLAG3) != 0;

    proc_t p1 = b1 ? Process1 : do_nothing;
    proc_t p2 = b2 ? Process2 : do_nothing;
    proc_t p3 = b3 ? Process3 : do_nothing;

    for (unsigned int i = 0; i<bytes; i++)
    {
        p1(data + i);
        p2(data + i);
        p3(data + i);
    }
}

int main()
{
    if (!QueryPerformanceFrequency(&ll)) return 1;

    const unsigned int bytes = 0xffff;
    srand((unsigned int)time(NULL));
    unsigned int flags = rand() & 0x7;
    unsigned char* data = new unsigned char[bytes];
    for (unsigned int i = 0; i < bytes; i++)
    {
        data[i] = (unsigned char)(rand() & 0xff);
    }

    START_CLOCK();

    ProcessData(data, bytes, flags);

    END_CLOCK();

    START_CLOCK();

    ProcessData_x(data, bytes, flags);

    END_CLOCK();
}

这是输出:

134 ns
272 ns

跑了好几次了，没想到，竟然耗费了更多的时间:(..还编译成'vs2010 Release x86'

Answer 1

首先，如果不考虑特定系统就谈论优化是没有任何意义的......

话虽如此，我将通过以下方式优化分支:

typedef void (*proc_t)(unsigned char*);

inline static void do_nothing (unsigned char* ptr)
{
    (void)ptr;
}

...

void ProcessData(unsigned char* data, unsigned int bytes, unsigned int flags)
{
    bool b1 = (flags & FLAG1) != 0;  // de-obfuscate the boolean logic
    bool b2 = (flags & FLAG2) != 0;
    bool b3 = (flags & FLAG3) != 0;

    proc_t p1 = b1 ? Process1 : do_nothing;
    proc_t p2 = b2 ? Process2 : do_nothing;
    proc_t p3 = b3 ? Process3 : do_nothing;

    for (unsigned int i = 0; i<bytes; i++)
    {
        p1(data + i);
        p2(data + i);
        p3(data + i);
    }
}