[Mlir-commits] [mlir] [mlir][vector] Don't fold in_bounds for scalable vector dimensions (PR #213506)

[Federico Bruzzone]

FedericoBruzzone wrote:

> @banach-space  Thanks for review. I don't have any SVE hardware, so I rented a `c7g.medium` (Graviton3, `ap-south-1`) for a few minutes specifically to check this one.
> 
> 
> 
> Method, in case it's useful: nothing was built on the instance. Both the folded and unfolded IR were cross-compiled locally via `mlir-opt` -> `mlir-translate` -> `llc -mtriple=aarch64-unknown-linux-gnu -mattr=+sve`, plus a small C driver that `mmap`s two pages, `mprotect`s the second `PROT_NONE`, and places the 4-element f32 buffer so its last element ends exactly at the boundary. The whole thing went in as a gzip+base64 EC2 user-data payload, so no key pair, security group or SSH was needed, results came back through `aws ec2 get-console-output`. Instance terminated on shutdown, cost about a cent.

Wow! Sorry to butt in, but this is really amazing! Thanks a lot for sharing these details. I think I'll try something similar in the future 🙏

Seriously, wow!

https://github.com/llvm/llvm-project/pull/213506