Dynamic MemRef Cast

[herhut]

@bondhugula

An extract_symbol operation has been missing since the beginning (has never been needed I think) to extract out the SSA values binding to the symbols of the memref’s affine layout map (these symbols are the same as strides and offset if the map is in the stride form). The dim op gets you the shape symbols but not the others, and you’ll need them for example to copy over the offset from the memref you are casting.
affine_map<(d0, d1) [s0, s1, s2] -> (s0 * d0 + s1 * d1 + s2)>

This is an interesting point. In the most general form, symbols can be referenced in the same way that the dimensions are, i.e., by providing an index. For the case of strided memrefs, I assume having some helper functions that identify the correct symbols for their meaning should be good enough.

So far this has been modeled only at the LLVM level I think, where we have access to these symbols as we can read them out of the memref descriptor.

I am all for moving ahead with this. We only need to bikeshed a name for the operation. get_symbol, extract_symbol? And the index as an attribute should be fine as this only makes sense in the ranked case.