Notes:
* Used DJGPP 2.01 and Allegro 3.1 final
* All results while running under Windows 95, no other apps running (so timer accuracy only 0.005).
* The tear time is the time for it to run the for loops with no code in the brackets.
* The system used was a P200MMX w/32 megs of ram and win 95

gcc -s -m486 -ffast-math -fomit-frame-pointer times:
Tear Time (for loop): 0.130
  Minus tear time: 0.000
Array access by [] operator: 0.530
  Minus tear time: 0.400
Array access by incrementing pointer: 0.535
  Minus tear time: 0.405

gcc -s -m486 -o3 -ffast-math -fomit-frame-pointer times:
Tear Time (for loop): 0.035
  Minus tear time: 0.000
Array access by [] operator: 0.230
  Minus tear time: 0.195
Array access by incrementing pointer: 0.225
  Minus tear time: 0.190

This is a segment of the code showing approximately what it is timing:
  tme = 0;
  install_int(timeproc, 5);
  for (int c=0; c<1000; c++) {
    arrptr = myarr;
    for (int d=0; d<5000; d++) {
      *(arrptr++) = d;
    }
  }
  remove_int(timeproc);

Conclusions:
  Even over 5,000,000 calculations, the difference between accessing by [] and accessing by pointer is immaterial in a sequential for loop.