Time consumption is different for 'for' loops with fix-length vector and dynamic vector

I have some test programs under the CCS3.3 and DSPC6713B and find that the time consumption is significantly different for the "for " loops.

 The first test program is as follows:

#define NUM   128
float dis[NUM];

#pragma MUST_ITERATE(10,,)
for(i = 0; i < NUM; i++)
   dis[i] =  dis[i] + val1 * val2; 

The second program is like:

typdef struct tag_OBJ
{
  float *dis;
}OBJ;

typdef struct tag_PARAM
{
   int num;
}PARAM;

OBJ obj;
PARAM param;

param.num = NUM;

obj.dis = (float *)malloc(sizeof(float) * param.num);

#pragma MUST_ITERATE(10,,)
for(i = 0; i < param.num; i++)
    obj.dis[i] = obj.dis[i] + val1 * val2;

The complie options for both cases is set to:-o3, -pm

It is surprised that time consumption for the first case is about 2-cycle for every loop and 20-cycle for each iteration for the second case?

What causes the dramatic difference on time consuption for these two cases?