[Mlir-commits] [mlir] [MLIR] Add a utility pass to linearize `memref` (PR #136797)

[Alan Li]

lialan wrote:

> But that's a bug because the index into the memref will be larger than the size
> 
> Please fix the PR so it that becomes a load from a memref<64xf32> instead
> 
> (My sense of the fix is that the getLimearizedDizeAndOffset utility should be doing `max_(d = 0 upto rank) {strides[d] * sizes[d]}` for non-identity layouts

Actually, I am not sure if that is a bug, because the way it is handled, there will be an `reinterpret_cast` to cast `memref<64xf132>` to `memref<16xf32>` before it gets fed into `memref.load`. 

I think this is more related to how we interpret "linearize indices", In my opinion this pass should only linearize `memref.load`'s indices but not change its memref size. We should call something like`FoldMemRefAliasOps` instead for folding memrefs.

https://github.com/llvm/llvm-project/pull/136797