There are some few cases where I'd probably prefer the "standard" version where available: memcpy etc. that are often extensively optimised, and that are a lot

There are some few cases where I'd probably prefer the "standard" version where available: memcpy etc. that are often extensively optimised, and that are a lot of work to write a non-naive version off that performs well. But I agree that most of the time you categorically should not use feature detection like that, and in the few cases you're dealing with systems lacking things that basic you have bigger problems...