Fix pipeline for view semantics

[tkarna]

• adds cse,canonicalize passes before one-shot-bufferize to remove duplicate bufferization.to_memref ops
  • if multiple identical bufferization.to_memref ops are present, one-shot-bufferize will create a copies
• adds a test that verifies that view semantics are correctly preserved

[fschlimb]

Nice.
How/why can duplicate to_memrefs happen?

[tkarna]

While adding cse does fix the view semantics by eliminating duplicate bufferization.to_memref ops, the elimination of common sub-expressions in general can change semantics. For example, in

a = sp.ones((6, 6))
b = sp.ones((6, 6))
c = a + b

the creation of b is removed as duplicate of a and a is used instead. Resulting in a double free (both a and b are still returned as memrefs).

[fschlimb]

Ok, so the duplicate free gets explicitly triggered by sharpy?

[tkarna]

Ok, so the duplicate free gets explicitly triggered by sharpy?

Yes, exactly. And if we want to keep the array api semantics, b in the above example should not be a reference to a.

[tkarna]

This issue is likely not manifesting in the new sharpy pipeline, closing.