GCC优化:如何减少运营速度?

时间:2016-05-06 07:53:04

标签: c++ gcc optimization

在尝试为我的代码测试某些选项时(使用128位整数)我观察到了一种我无法理解的行为。有人可以对此有所了解吗?

#include <stdio.h>
#include <stdint.h>
#include <time.h>

int main(int a, char** b)
{
    printf("Running tests\n");

    clock_t start = clock();
    unsigned __int128 t = 13;
    for(unsigned long i = 0; i < (1UL<<30); i++)
        t += 23442*t + 25;
    if(t == 0) printf("0\n");
    printf("u128, +25, took %fs\n", double(clock() - start)/CLOCKS_PER_SEC);

    start = clock();
    t = 13;
    for(unsigned long i = 0; i < (1UL<<30); i++)
        t += 23442*t;
    if(t == 0) printf("0\n");
    printf("u128, no+, took %fs\n", double(clock() - start)/CLOCKS_PER_SEC);

    start = clock();
    unsigned long u = 13;
    for(unsigned long i = 0; i < (1UL<<30); i++)
        u += 23442*u + 25;
    if(u == 0) printf("0\n");
    printf("u64 , +25, took %fs\n", double(clock() - start)/CLOCKS_PER_SEC);

    start = clock();
    u = 13;
    for(unsigned long i = 0; i < (1UL<<30); i++)
        u += 23442*u;
    if(u == 0) printf("0\n");
    printf("u64 , no+, took %fs\n", double(clock() - start)/CLOCKS_PER_SEC);

    return 0;
}

(请注意,printf在这里,以便gcc不会优化for循环) 在我的系统上,这可靠地产生以下输出:

u128, +25, took 2.411922s
u128, no+, took 1.799805s
u64 , +25, took 1.797960s
u64 , no+, took 2.454104s

虽然128位整数行为是有意义的,但我没有看到具有较少操作的64位循环如何显着(30%)较慢。

这是一种已知行为吗?在编写此类循环时尝试从此优化中受益的一般规则是什么?

编辑:仅在使用-O3选项进行编译时才会观察到该行为。

gcc -lstdc++ -O3 -o a main.cpp

u128, +25, took 2.413949s
u128, no+, took 1.799469s
u64 , +25, took 1.798278s
u64 , no+, took 2.453414s

gcc -lstdc++ -O2 -o a main.cpp

u128, +25, took 2.415244s
u128, no+, took 1.800499s
u64 , +25, took 1.798699s
u64 , no+, took 1.348133s

2 个答案:

答案 0 :(得分:8)

循环太紧,依赖性失速,ALU忙等起作用并控制时机。因此,与实际指令执行相比,结果不可靠且对其他因素更敏感。

请注意,+ 25可以与乘法并行计算。

PS。我在4970K上的结果:

gcc version 5.2.1 20151010
gcc -lstdc++ -O2 -o a a.cpp

u128, +25, took 1.346360s
u128, no+, took 1.022965s
u64 , +25, took 1.020189s
u64 , no+, took 0.765725s

编辑:在查看-O2-O3上的反汇编后,主要区别在于代码生成。 (由于上述原因,在不同的测试机器/环境中仍然保持-O2,产生的结果略有不同)

<强> -O2:

400618:       48 69 d2 93 5b 00 00    imul   $0x5b93,%rdx,%rdx
40061f:       48 83 e8 01             sub    $0x1,%rax
400623:       75 f3                   jne    400618 <_Z4testv+0x18>

<强> -O3:

400628:       66 0f 6f d9             movdqa %xmm1,%xmm3
40062c:       83 c0 01                add    $0x1,%eax
40062f:       66 0f 6f c1             movdqa %xmm1,%xmm0
400633:       66 0f f4 cc             pmuludq %xmm4,%xmm1
400637:       3d 00 00 00 20          cmp    $0x20000000,%eax
40063c:       66 0f f4 da             pmuludq %xmm2,%xmm3
400640:       66 0f 73 d0 20          psrlq  $0x20,%xmm0
....

O3生成矢量化代码,而循环具有很强的依赖性,无法从矢量化中获取值。它实际上产生了更复杂的代码,因此具有更长的时间。

答案 1 :(得分:0)

您需要测量运行时实际发生的事情才能确定。 正如Calvin中提到his answer一样,处理器内部会发生很多事情,这会影响循环的最终时间。

您可以使用PAPIIntel's brilliant tools来执行测量。英特尔的工具价格昂贵,但您可以免费试用30天。它们比PAPI更容易使用。