[llvm] [SystemZ] Enable MachineCombiner for FP reassociation (PR #83546)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

> A bit surprisingly, I saw instead a 13% improvement imagick _without_ the MachineCombiner, just because of the global reg/mem folding I tried. 

Haven't looked into the patch in detail yet, but this is interesting.  Should we then always do the reg/mem folding late (for all FP operations, including non-reassociable ones), if it is generally beneficial?  This might actually even simplify the code (and could be done as a separate PR before the MachineCombiner reassociation patch?).

https://github.com/llvm/llvm-project/pull/83546