C++ 用GCC实现自动矢量化

我想在我的代码中矢量化矩阵向量积。我尝试在GCC中使用自动矢量化，但它根本不起作用，我也不知道如何使它起作用。现在我尝试一个非常简单的示例代码：

#define N 200000
double a[N] __attribute__(( aligned(16) ));
double b[N] __attribute__(( aligned(16) )) ;
double c[N] __attribute__(( aligned(16) )) ;

void f()
{ 
    int i;

    for( i = 0; i < N; i++ )
    {
        a[i] = b[i] + c[i];
    }
}

int main( int argc,char** argv )
{
    f();

    return( 0 );
}

我得到了以下结果我真的不明白这些东西是什么意思，但底线显然是矢量化不起作用：

optimization2.cpp:10:2: note: misalign = 0 bytes of ref b[i_11]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref c[i_11]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref a[i_11]
optimization2.cpp:10:2: note: virtual phi. skip.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:6:6: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:12:13: note: not vectorized: no vectype for stmt: vect__4.6_1 = MEM[(int *)vectp_b.4_9];
 scalar_type: vector(4) int
optimization2.cpp:12:13: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:6:6: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:14:1: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:10:2: note: misalign = 0 bytes of ref b[i_12]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref c[i_12]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref a[i_12]
optimization2.cpp:10:2: note: virtual phi. skip.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:12:13: note: not vectorized: no vectype for stmt: vect__4.28_2 = MEM[(int *)vectp_b.26_9];
 scalar_type: vector(4) int
optimization2.cpp:12:13: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.

---

我尝试过很多其他的东西和代码，但从来没有得到任何矢量化。诀窍是什么？

当我使用gcc 4.9.2和相同的命令行编译它时，虽然-march=native可能意味着不同的东西，添加-save temp，但我从优化器和包含vmodapd和vaddpd的程序集获得相同的输出，即向量化。我认为优化器注释是关于代码中无法矢量化的其他部分。您确定生成的asm中添加的内容没有矢量化吗？@Wintermute：从其他示例中，我的印象是GCC显式声明循环是否已矢量化详细级别不会改变这一点。代码由我发布的内容组成，没有其他循环会导致矢量化失败。我还没有看过汇编代码。但是，在假设-O2的情况下，无论是否激活矢量化，似乎都不会有速度增益。没有/几乎没有性能增益是因为gcc 4.9.2生成SSE指令，尽管即使使用-O0也不会生成SSE2。SSE2指令仅使用-O3生成，但在这种情况下它们比SSE1的优势是有限的。为什么优化器输出没有提到任何这一点是任何人的猜测，但我正在查看生成的asm，在那里它是清晰可见的。

optimization2.cpp:10:2: note: misalign = 0 bytes of ref b[i_11]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref c[i_11]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref a[i_11]
optimization2.cpp:10:2: note: virtual phi. skip.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:6:6: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:12:13: note: not vectorized: no vectype for stmt: vect__4.6_1 = MEM[(int *)vectp_b.4_9];
 scalar_type: vector(4) int
optimization2.cpp:12:13: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:6:6: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:14:1: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:10:2: note: misalign = 0 bytes of ref b[i_12]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref c[i_12]
optimization2.cpp:10:2: note: misalign = 0 bytes of ref a[i_12]
optimization2.cpp:10:2: note: virtual phi. skip.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:10:2: note: num. args = 4 (not unary/binary/ternary op).
optimization2.cpp:10:2: note: not ssa-name.
optimization2.cpp:10:2: note: use not simple.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:12:13: note: not vectorized: no vectype for stmt: vect__4.28_2 = MEM[(int *)vectp_b.26_9];
 scalar_type: vector(4) int
optimization2.cpp:12:13: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.
optimization2.cpp:16:5: note: not vectorized: not enough data-refs in basic block.