[Mlir-commits] [mlir] [MLIR][XeGPU] Add anchor_layout and update propagation to honor user-specified layouts (PR #169267)

[Sang Ik Lee]

silee2 wrote:

Can you explain with load, store, prefetch is an anchor op?
My understanding of anchor op is, it is some kind of a co-operative op and requires a specific layout.
Load/store/prefetch is regular SIMT instruction. They just appear as vector operation at WG level but becomes SIMT after SG LANE distribution.
As such I don't think those ops will actively demand a certain layout.
Aren't those ops playing passive roles for layout decision?

https://github.com/llvm/llvm-project/pull/169267