[flang][hlfir] Better recognize non-overlapping array sections.

[vzakhari]

This is a copy of the corresponding ArrayValueCopy analysis
for non-overlapping array slices. It is required to achieve
the same performance for Polyhedron/nf, though, additional
changes are needed in the alias analysis for disambiguating
host associated accesses.

[jeanPerier]

Looks great to me

[tblah]

Looks good to me! Nice work. Is that all of the performance stuff done?

[vzakhari]

Looks good to me! Nice work. Is that all of the performance stuff done?

Unfortunately, this does not resolve the nf performance regression. In the benchmark, in addition to non-overlapping slices of x on the LHS and RHS, there is a use of a host associated variable on the RHS, and this use is reported as conflicting by the current alias analysis. ArrayValueCopy is able to claim no-overlap in this case somehow. The chain of operations for this host associated variable access up to the dummy block argument (for the host associated variables) looks quite complex, e.g. there are indexing of the host associate block and box loads. I think we may need to attribute the corresponding hlfir.declare with fir.host_assoc so that the alias analysis can disambiguate dummy argument accesses vs the host associated variables accesses.