As discussed on llvm-dev recently, we should refine our FP transforms based on the minimum subset of fast-math-flags needed to enable those transforms.
These intrinsic folds were just added with D41381, but we let them slide with isFast(). I think 'reassoc' is enough given that we're just eliminating ops based on their math inverses.
'reassoc' instead of 'fast' should also be all that's needed for the llvm.log.f64 call to enable this transformation.