This experiment studies the speed-up increase that alias analysis (AA) produces on code for very long instruction word machines. AA is done on-demand when requested by the scheduler, in order to eliminate critical arcs of the data dependence graph. Different heuristic criteria are investigated for deciding when to compute alias information,and they show that only a fraction of the alias relation really contributes to the program speed-up. A qualitative study shows that the quality of the initial code affects the speedup alias analysis can give. The results should help compiler designers for VLIW machines in making cost effective AA decisions.