Limit calculation of pre loop during super word optimization is wrong

gunter.haug@sap.com reported the following problem:

The variable "MemNode* best_align_to_mem_ref" is hidden by a second declaration of a variable with the same name in an inner block of the function "SuperWord::find_adjacent_refs()". This leads to unaligned accesses in generated code. While many platforms handle these correctly (slower, though), Itanium and potentially other platforms which don't allow unaligned access will crash.