Look at __builtin_fmax() and fmin() again. They seem to be slower than a simpler ternary. See #1729