[llvm] [AMDGPU] DAG Mutation to solve load bunching in outer product matrix multiplications (PR #203095)

[Axel Sorenson via llvm-commits]

================

----------------
axelcool1234 wrote:

Good catch. Instead of min/max, I went a little further and directly used each load's latency and reciprocal throughput when needed.

https://github.com/llvm/llvm-project/pull/203095