[llvm] [AArch64][SelectionDAG] Fold late FCMP+CSEL patterns to FMAXNM/FMINNM during AArch64 DAG-to-DAG preprocessing. (PR #197429)

[Jacques Pienaar via llvm-commits]

jpienaar wrote:

Thanks! This was extracted from XLA compiling bf16 tanh, in case helpful I can dump the LLVM IR (but seeing as you are seeing potential fix, I could perhaps help test it as easier way)

https://github.com/llvm/llvm-project/pull/197429