13:12 <tos9> oy, it's 9AM and I bet I just found a threading bug in PyPy sqlite

13:24 <tos9> *phew*, ok, that was easier than expected, it's not a threading bug, it's a pypy3 bug

13:29 <tos9> (bug is https://foss.heptapod.net/pypy/pypy/issues/3226 )

14:13 <cfbolz> tos9: that's fixed in py3.7, might need a backport

14:14 <tos9> cfbolz: a ha! cool, I scratched my head for 5 mins trying to figure out how this can be broken because tests in the sqlite suite seem to cover it, but I was looking at cpython3.8

14:14 <cfbolz> 9c27af1d34e9

14:15 <tos9> seems to already be grafted

14:15 <tos9> maybe just not released

14:16 <cfbolz> right

14:18 <tos9> thanks!

14:18 <cfbolz> tos9: is it fixed on py2.7 too?

14:18 <tos9> cfbolz: in PyPy2 it works for me already

14:18 <cfbolz> ah, ok

14:18 <tos9> (in 7.3.1)

14:29 <marmoute> ho, I as mistaken then :-)

14:29 <marmoute> looks at your changes now then

20:40 <mattip> cfbolz: could you unpin the top tweet on pypyproject?

23:12 <catern> haha, sob, I went to all this effort to get my python-3.7-requiring stuff running with pypy and... my benchmark is 50% slower

23:12 <catern> or more like 80% slower

23:19 <catern> hm, I did build from source - is there a quick way to benchmark pypy to check that I didn't, maybe, build without optimizations?

23:21 <tos9> you have to turn that off during the build with IIRC --no-jit or something like that

23:21 <tos9> did someone say that wasn't expected though?

23:22 <tos9> the py3.7 branch is AIUI not near ready

23:22 <ronan> performance should be similar to 3.6 though

23:22 <tos9> I don't know that that's in ways beyond the stdlib (and IIRC some parser changes? I didn't catch the scrollback entirely), but other things being missing is definitely not out of the question I'd suspect.

23:22 <tos9> ah ok so now ronan confirme dit :)

23:22 <tos9> never mind me

23:23 <ronan> catern: can you share your benchmark?

23:29 <catern> hmm it's not totally trivial, it uses this custom C library which I only build with Nix right now https://github.com/catern/rsyscall

23:32 <catern> i noticed a big swing in performance of a couple functions in my profile, so I'm looking at that right now

23:32 <catern> just to clear it up, are coroutines expected to perform well? since this is all async code

23:38 <catern> maybe this is all trio's fault, I'm noticing a lot of slowdowns in trio primitives, maybe it does something weird...

23:48 <catern> (hmm no, false alarm on trio primitives, it seems to be something weirder; cprofile is reporting a lot more calls to a bunch of my polling functions, maybe that's measurement error, but maybe it's some weird scheduling thing)

23:49 <catern> (if I just divide total time by number of calls, individual functions do seem to have sped up a lot at least :))

23:55 <catern> I also ran a simpler benchmark without cprofile and got roughly the same slowdown so I don't think cprofile is at fault