Why using scaled offsets in block ptr?

[Nullkooland]

I'm working on a solution for #203, and I noticed that there's a difference of offsets param of tt.make_tensor_ptr and tts.make_tptr. tt.make_tensor_ptr is using per-dimension offests while tts.make_tptr is using scaled offsets which is multiplied by strides. So during conversion of tt.make_tensor_ptr to tts.make_tptr, the per-dimension offests is converted to scaled offsets by inserting arith.muli op with the strides (See #90), however I need the per-dimension offsets to figure out the N-D masked shape on the last block boundary.

I wonder what's the rationale behind adopting the scaled offsets early in tts.make_tptr, since the strides param is also included and we can get the scaled offsets later in structured-to-memref pass.

@nhat-nguyen @haishanzzzz Could you take a look?

[nhat-nguyen]

Thanks for your question! This is an interesting observation. I'll take a look and get back to you with some thoughts.