A point: Benchmarks produce metrics measured in seconds. Bugs don't produce metrics you can measure but the costs are unbounded.
Aren't you contradicting yourself here?
And I do think that some UB optimizations make a lot of sense. Actually if you look at the way Rust does it, it gets rid of many UBs not by adding checks for them but by explicitly forbidding the UB condition to occur in safe code (therefore "optimizing on UB" by enforcing that the UB condition cannot happen in correct code, not by adding checks for it and slowing things down). That's why for instance pointer arithmetic is unsafe in Rust, since arbitrary pointer offsets can lead to UB.
Although I agree that there are a few C UBs that I think ought to be relaxed because they don't make a lot of sense for modern architectures though (signed overflow being the main one).
Bringing up signed overflow. I did some tests once flipping back and forth between using a int and an unsigned int and wasn't impressed with the speed up. I've also done tests with aligned and an packed data structures and on a modern processor, also not impressed.
I don't think it's a contradiction. I think what's being said is that it's the wrong default, i.e. the special different-looking operators should be the ones with UB, not the obvious "normal"-looking ones.