[RFC] What is the intended semantic of `memref.reinterpret_cast`?

[ftynse]

@qcolombet

Is my understanding correct that reinterpret_casts are noops with respect to base + offset? (I.e., if %offset is not equal to src.offset at runtime, then all bets are off), or

This is also my understanding. reinterpret_cast can be used to add or remove shape or rank dynamicity to memrefs, but nothing else. This is also what the verifier checks when it can https://github.com/llvm/llvm-project/blob/142aa1bdd1dd1db9a7fecf9d157228019c794c94/mlir/lib/Dialect/MemRef/IR/MemRefOps.cpp#L1839-L1879.

@qcolombet

I don’t think that’s the intended semantic of memref.reinterpret_cast though, i.e., I believe that technically we could lower memref.reinterpret_cast with some runtime assertions (who’s up to build a mlir-ubsan? :))

Isn’t this done by [RFC] Runtime Op Verification ?

@qcolombet

I’ll post another RFC with a proposed fix to re-building a memref from the results of memref.extract_strided_metadata operations.

Having such an op gets quite scary. It allows one to effectively break out of the memref structure and create arbitrary, potentially invalid memrefs. Most memref-based analyses become invalid because they can no longer assume memref well-formedness in presence of such an op.