[Mlir-commits] [mlir] [mlir] Add direct vectorization lowering for `tensor.pack` ops (PR #78660)

[Diego Caballero]

dcaballe wrote:

Ok, I'm anticipating a problem here that I'm hitting right now: the direct vectorization pattern has to generate good code when the target doesn't support masking (e.g., Arm Neon). Currently, if we try to vectorize a tensor.pad op without masking, we generate transfer reads with the `in_bounds` flags set to `false`, which causes many other problems down the road (e.g., some other canonicalization patterns not triggering). We have to make sure that whatever direct ops we generate, we have an good path for targets without a mask. I haven't thought too much about this but perhaps we can try to apply peeling to the pad op when masking is not supported.

https://github.com/llvm/llvm-project/pull/78660