GEMM performance regressed after sinking unused l-invariant loads

I was looking into some performance regression cases in our downstream application and found https://github.com/llvm/llvm-project/pull/157559 changed the generated llvm codes and thus affect the performance.

I'm attaching the before and after .ll file for reference.
https://gist.github.com/yzhang93/f7174c373ec63845d63762997b3ee293

Start by comparing the before and after .ll files in the linked gist, then inspect pull request 157559 to identify the sinking change. Reproduce the downstream GEMM case if available and measure the generated-code performance; done means addressing the regression without losing the intended optimization.

Written by the indexing model from the issue text.