Mill Computing, Inc. Forums The Mill Architecture Support for multidimensional arrays Reply To: Support for multidimensional arrays

Ivan Godard
Post count: 689

These are commonly called “dope vectors” and are used in Fortran and Ada among others. Any single dimension is a normal indexing access (effectively a LEA for all dimensions but the last). The multidimensional case is hard to put into the hardware because it needs a way to encode several indices, and most hardware has hard limits on the number of arguments per op (not the Mill though). Another issue is the size of the dope vector itself, needing a stride and offset at each level, which gets out of hand after the second dimension. However, the biggest reason for not directly supporting dope vectors is the embedded stride multiply: you can overlap all the rest of the indexing with the multiplies, so making it one op rather than several doesn’t buy anything.