I see a factor of 2 in the closest benchmarks. A factor of 4-5 for the middlish ones, and a factor of 10 for the least close. But it's worth noting that a lot of the those C++ examples are making heavy use of SIMD.
I kinda got the feeling from C/C++ compilers that auto-vectorization is brittle and unpredictable even there and that's with all the time an AoT compiler can invest into optimization. A JIT always has to tread the line (or multiple lines in case of tiered compilation) between optimization effort and benefit. And the usual trade-off between larger and smaller code. So speculatively unrolling short loops just to see whether anything could be vectorized is perhaps out of scope for most JITs and even in AoT-compiled languages the preferred way of getting the most out of vector instructions is just to explicitly vectorize hot loops as a programmer, as the result is often better than what the compiler could do.