BrowserUK: Thanks for the code examples and ideas. For me this is also a nice example of using Inline (the first time for me). I only had to change unsigned __int64 bits in code>diffBits</code> which my toolchain (Mac OSX) can not handle.
The problem with that benchmark is it only tests the call to the functions; not the subsequent accessing of the returned information.
The benchmarks of diff implementations mentioned in my blog return all the same: a reference to an AoA. If they return the same, I need not care about subsequent accessing. The exception is String::Similarity with (fast) strings as input, without constructing the aligment, and returning only one number. It's only there for showing an upper limit