On Friday, February 26, 2016 09:48:38 AM Kristoffer Carlsson wrote: > It is possible that micro benchmarks might be misleading since in real code > you might trash the instruction cache if you load a fat unrolled monster > function into it.
The nice thing is that the compiler knows what K is and will generate different code for each choice of K---if LLVM is smart about its unrolling, presumably it will unroll for small K but not for large K. Of course, it would be even nicer if you could say "use the default fallback for any K > 8". But I don't think we can do that right now. Best, --Tim
