That is not the same test as the OP's. It's not even the same test between the two different algorithms, since you hide different information from LLVM under d

That is not the same test as the OP's. It's not even the same test between the two different algorithms, since you hide different information from LLVM under different circumstances. If you have to put `test::black_box` everywhere to get anything but zeroes, the only thing you can really conclude is that LLVM is better at optimizing than you are at writing microbenchmarks (I'll agree it can be frustrating at times).