Alignment not working properly on my sse program

Question

I'm having a problem with my little program as follows:

#include <iostream>
#include <xmmintrin.h>
using namespace std;
int Main()
{
  __m128 mm_sum;
  __m128 mm_buf;

  float array1[4] __attribute__ ((aligned(16)));
  array1[0] = 1.1;
  array1[1] = 2.1;
  array1[2] = 3.1;
  array1[3] = 4.1;

  mm_buf = _mm_load_ps(array1);
  mm_sum = _mm_add_ps(mm_sum, mm_buf);
  float *p1 = new float[4];
  _mm_store_ps(p1,mm_sum);

  float Sum = p1[0]+p1[1]+p1[2]+p1[3];
  cout<<"Sum= "<<Sum<<endl;

  delete[] array1;
  delete[] p1;
  return 0;
}

I'm using gcc 4.8.1. It appears that "attribute((aligned(16)))" doesn't work properly because when I repeat compiling and running the program, it suddenly gives me an odd value like -1.759e+20 after every four or five time of correct result. I checked other posts but didn't find a good solution. Please help~!! Thanks a lot!!

Answer 1

You have a problem here:

float *p1 = new float[4];
_mm_store_ps(p1,mm_sum);

p1 is not guaranteed to be 16 byte aligned, so _mm_store_ps may fail. Either ensure that p1 is 16 byte aligned, or use _mm_storeu_ps instead of _mm_store_ps.

You have another problem here:

delete[] array1;

array1 was not allocated via new [] so this will result in Undefined Behaviour.

Also mm_sum needs to be initialised.

Here is a complete working version of your code:

#include <iostream>
#include <xmmintrin.h>

using namespace std;

int main()
{
    __m128 mm_sum = _mm_setzero_ps();   // <<< mm_sum needs to be initialised
    __m128 mm_buf;

    float array1[4] __attribute__ ((aligned(16)));
    array1[0] = 1.1;
    array1[1] = 2.1;
    array1[2] = 3.1;
    array1[3] = 4.1;

    mm_buf = _mm_load_ps(array1);
    mm_sum = _mm_add_ps(mm_sum, mm_buf);
    float *p1 = new float[4];
    _mm_storeu_ps(p1,mm_sum);           // <<< use unaligned store as p1 is not aligned

    float Sum = p1[0]+p1[1]+p1[2]+p1[3];
    cout<<"Sum= "<<Sum<<endl;

    delete[] p1;
                                        // <<< NB: do NOT delete [] array1 !
    return 0;
}

Test it:

$ g++ -Wall -msse3 so_storeu_ps.cpp && ./a.out
Sum= 10.4
$