Possible regression on LSR

Hello,

I believe I found a possible regression introduced by #184138. The regression shows up in matrix matrix multiplication. I have some high level benchmarks that track the performance of matrix multiplication for two matrices with of size 8192x8192 and with f32 values.

After PR #184138, I see a 1.81x slowdown in multiplying two matrices of size 8192x8192 for gfx1201.

I am attaching the output of `llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1201 -print-before=loop-reduce -print-after=loop-reduce` for both, what exists on main (referencing 76cdf6012b883529e3d608ed4e30d622ba43e10c) and applying a revert of #184138.

[main.txt](https://github.com/user-attachments/files/27730265/main.txt)

[revert.txt](https://github.com/user-attachments/files/27730262/revert.txt)