[llvm] [AMDGPU] Don't cluster DS-instrs together (PR #180908)

[via llvm-commits]

lijinpei-amd wrote:

> Can you add more details about the motivation for the PR and what it does?

Sure.

The motavation:

We were analyzing some assembly generated by triton/gluon matmul kernels, and found that if ds_read instrs were not scheduled together, we got 10% perf increase for our case. Further analysis shows that ds_reads are scheduled together because of schedule-dag mutation and SIPostRABundler.

What it does:

The goal of this pr is to remove the cluster-edge created by dag mutation, and remove the bundle created by SIPostRABundler, so that scheduler has more freedom to move lds instrs around.

I have updated the pr description and uploaded the llvm-ir and assembly of my cases here: https://gist.github.com/lijinpei-amd/0d114c50891abe5bebb564e2084d357d .

https://github.com/llvm/llvm-project/pull/180908