These are chat archives for dropbox/pyston

1st
May 2015
Michael Arntzenius
@rntz
May 01 2015 00:38
if I change libunwind, how do I tell cmake that I've done it?
it doesn't seem to auto-detect it
Daniel Agar
@dagar
May 01 2015 00:38
delete it in the build directory
Chris Toshok
@toshok
May 01 2015 01:34
wow, finally
sums      = 18086025
main time = 18887744
:sparkles:
Marius Wachtler
@undingen
May 01 2015 13:11
awesome :+1:
Travis Hance
@tjhance
May 01 2015 15:55
What is that?
Chris Toshok
@toshok
May 01 2015 15:55
the nested timer stuff. up until last night the sums was greater than the main time (a timer wrapping the body of main)
it’s still not the best that they aren’t closer - they’re in us, so they’re ~800ms apart, but at least they’re 800ms apart in the right direction
is there some better way to get the cpu frequency on linux other than reading /proc/cpuinfo?
the kernel obviously knows it, since that’s what populates /proc
Chris Toshok
@toshok
May 01 2015 17:39
ugh:
toshok@jitdev:~/pyston$ ./pyston_dbg test/tests/weakref4.py
3628800
None
toshok@jitdev:~/pyston$ ./pyston_release test/tests/weakref4.py
3628800
function
synecdoche
@synecdoche
May 01 2015 18:15
@toshok if you want to account for cpu scaling, maybe this? /sys/devices/system/cpu/cpu${N}/cpufreq/cpuinfo_cur_freq
Chris Toshok
@toshok
May 01 2015 18:16
scaling shouldn’t be an issue - apparently newer intels/amds have a constant tsc
which is all kinds of awesome
synecdoche
@synecdoche
May 01 2015 18:16
oh yeah, right
and tscs should even be in sync if the cpus are on the same chip :)
Chris Toshok
@toshok
May 01 2015 19:21
not sure how the weakref tests are flapping on my PR, since I do nothing GC related at all there :/
Kevin Modzelewski
@kmod
May 01 2015 20:43
well, those tests have to make sure that the object gets GCd
I think if it says "function" at the end that means that it didn't GC the object
Chris Toshok
@toshok
May 01 2015 20:43
right
Kevin Modzelewski
@kmod
May 01 2015 20:43
so I think it's kind of sensitive to stack layout and what not
Chris Toshok
@toshok
May 01 2015 20:44
ohh, that makes some sense, since I think the PR has the timers enabled
Kevin Modzelewski
@kmod
May 01 2015 20:44
maybe we could have something like __pyston__.zero_stack() which overwrites a bunch of the stack
one idea to make it more robust is to do the work farther down the stack
so the references to the object we want to GC are less likely to be in the current stack range at the time of the GC?
Chris Toshok
@toshok
May 01 2015 21:24
i could also see about making the StatTimer’s a little thinner
right now they’re 28 bytes
Kevin Modzelewski
@kmod
May 01 2015 21:26
that doesn't seem like much :P
Chris Toshok
@toshok
May 01 2015 21:27
yeah, I thought it’d be more. it was yesterday… I still had a Timer in there, not doing anything at all
Chris Toshok
@toshok
May 01 2015 21:51
yup, disabling slowpath timers fixes the gc tests