I have to say that from the document I get the feeling that clang makes the best compromise; their implementation seems much more clear and linear to me, and it

I have to say that from the document I get the feeling that clang makes the best compromise; their implementation seems much more clear and linear to me, and it looses just on character in short mode. Benchmarking these kind of things is always difficult, because you can rig (voluntarily or not) microbenchmarks to go either way, and the performance gains are easily lost in real world usage.