Re: LuaJIT2 performance for number crunching

[Mike Pall <mikelu-1102@...>]

Francesco Abbate wrote:
> Do you think this problem could be eventually fixed in LuaJIT2?

The loop unrolling heuristics decided that the loop should be
unrolled, but then discovered that the loop body was too long and
gave up (instead of falling back to a non-unrolled loop). Sigh. :-/

I guess I have to completely redesign the loop unrolling heuristics.
This will take me quite a bit of time, because I need to verify that
it still performs well for many different benchmarks. Alas, other
work has priority right now, so don't hold your breath.

--Mike