[llvm] [X86] Prefer SIMD min/max/abs for scalars when staying in XMM domain (PR #210654)

[Simon Pilgrim via llvm-commits]

https://github.com/RKSimon commented:

If you're matching against a RMW pattern - why not begin the fold from combineStore - similar to narrowBitOpRMW?

https://github.com/llvm/llvm-project/pull/210654