[llvm] [AMDGPU][DAGCombiner][GlobalISel] Prevent FMA contraction when multiply cannot be eliminated (PR #169735)

[via llvm-commits]

b-sumner wrote:

I don't think the FMA use in our log implementation reduced the overall critical path length.  Is any guarantee being made that it will when applied?  Also, is any consideration being given to power use?  On some implementations a floating point add uses less power than FMA.  Also on some implementations floating point add uses fewer resources for less time than FMA.

https://github.com/llvm/llvm-project/pull/169735