[Mlir-commits] [mlir] [mlir][affine] Select shared-input reductions for unroll-and-jam (PR #226270)

[llvmlistbot at llvm.org]

ParkHanbum wrote:

For additional context, closely related ideas have been discussed in prior work on register blocking / unroll-and-jam for convolution-like loop nests, where unrolling outer loops is used to expose reuse of shared inputs in the inner reduction.

This PR is intentionally much narrower. It only applies unroll-and-jam to a specific, structurally verified pattern where multiple independent reductions share the same input, and only after the required legality conditions are proven.

So the goal here is not to implement the broader optimization strategy from the paper, but to capture one concrete and well-defined case of the same underlying reuse opportunity in Affine.

Related work:
Anthimopoulos et al., “Register Blocking: A Source-to-Source Analytical Modelling Approach for Affine Loop Kernels,” ACM TECS, 2025.

https://github.com/llvm/llvm-project/pull/226270