A loop didn’t have to build a stack frame in the first place. Once the optimizer has done it’s job, both the TCO optimized tail call and a basic loop will have the same instructions, hence the performance will be the same.

Yes, that's exactly the point I am making.

Modern compilers optimise this for you, exactly as you said.

Back in the 80s compilers didn't do this. There really was a difference between the machine code emitted for a naive loop and a tail call loop.

A lot of advice to rewrite naive loops to do tail calls is based on that 1980s compiler behaviour, not on the optimising 2020s compiler behaviour.