These are chat archives for dropbox/pyston

Mar 2016
Kevin Modzelewski
Mar 25 2016 03:11
So I spent today trying to understand refcounting performance
there are some pretty significant regressions in the amount of time it takes to run jitted code
at first I thought it was from the added refcounting instructions
but I think it's actually from general icache / process frontend contention
Even an empty function (with basically no refcounting operations) gets about 20% slower
Kevin Modzelewski
Mar 25 2016 03:17
My current theory is that it's because we're interspersing deallocation code through everything
But anyway, I don't really know what we can do about the general issue
There are a couple more optimizations that we can do to refcounting performance so hopefully that will bring it back to parity
Marius Wachtler
Mar 25 2016 14:40
Hi, I could not yet finish all deopt problems (still leaking stuff), but I will try to continue to work on it later today. But I will probably not have internet until tomorrow