These are chat archives for dropbox/pyston

16th
Jul 2015
Marius Wachtler
@undingen
Jul 16 2015 00:05
does anybody else see the pyicu test sometimes failing on travis-ci?
Chris Toshok
@toshok
Jul 16 2015 00:05
i’ve had it happen
Marius Wachtler
@undingen
Jul 16 2015 00:06
oh cool :-) (the video not the error!)
Rudi Chen
@rudi-c
Jul 16 2015 00:09
Is jitdev down for anyone else?
Daniel Agar
@dagar
Jul 16 2015 00:11
Just adding -flto to the gcc build seems to work
Chris Toshok
@toshok
Jul 16 2015 01:11
Not globally I'm guessing?
Kevin Modzelewski
@kmod
Jul 16 2015 01:15
also, sorry for misleading you; I don't think we want to use the gold plugin from our llvm build, we want to use the one from whatever clang we used to compile our code
I think as long as we pass in the already-compiled stdlib.bc.o file (or whatever its name is) we should be ok
I'm assuming it can handle a mix of native object files and bitcode
I think just copying what cpython does though is probably the easiest way to start
-flto -fprofile-generate then -flto -fprofile-use
Daniel Agar
@dagar
Jul 16 2015 01:20
@toshok yes globally
Chris Toshok
@toshok
Jul 16 2015 01:20
The gold plugin from the system clang was crashing whereas the pyston built one wasn't. I suspect that the crash was due to the same issue that caused the warnings/errors with the custom built llvmgold.so, but don't know for sure. It does seem to deal fine with object + bit code, which is pretty awesome.
Daniel Agar
@dagar
Jul 16 2015 01:21
ha, and it's slower
Chris Toshok
@toshok
Jul 16 2015 01:21
That's strange - clang with -flto will generate bit code in the libstdlib.a
Also, stdlib.bc.o was a normal object. I don't think that file was the cause of the problems.
Kevin Modzelewski
@kmod
Jul 16 2015 01:24
are you using the system clang on your jitdev? I think our instructions say to install a different one
Chris Toshok
@toshok
Jul 16 2015 01:26
Hm, at least for the link line it was /usr/bin/clang++. I didn't verify it was for the individual compiles.
Chris Toshok
@toshok
Jul 16 2015 19:51
<jasone> toshok, here's the result of running the pyston benchmark: https://github.com/jemalloc/jemalloc/commit/aa2826621e1793db9faea31e803690ccbe36f14c and https://github.com/jemalloc/jemalloc/commit/5bd879646cdb303045096fc2b9d7
Chris Toshok
@toshok
Jul 16 2015 20:28
still, not much of a reason to switch imo:
                               jemalloc-dev~1:        jemalloc-dev:
       django_template.py             3.6s (2)             3.6s (4)  -0.4%
            pyxl_bench.py             3.5s (2)             3.5s (4)  +0.4%
sqlalchemy_imperative2.py             4.2s (2)             4.2s (4)  +0.1%
        django_migrate.py             1.5s (2)             1.5s (4)  +1.1%
      virtualenv_bench.py             4.8s (2)             4.9s (4)  +0.3%
                  geomean                 3.3s                 3.3s  +0.3%
Chris Toshok
@toshok
Jul 16 2015 20:58
:+1: investigate.py resolves the symbol references to their hashes and caches them
Chris Toshok
@toshok
Jul 16 2015 22:39
has anyone else been experiencing test timeouts?
on travis-ci, that is
Rudi Chen
@rudi-c
Jul 16 2015 22:40
If it's CPython tests I increased the timeout in my finalizer diff.
Chris Toshok
@toshok
Jul 16 2015 22:43
yeah, I ran into it as well yesterday, just odd to have it start happening
Chris Toshok
@toshok
Jul 16 2015 22:54
oh funky, since the different PR’s share the same base commits, comments on the commits show up on all the PR's
Rudi Chen
@rudi-c
Jul 16 2015 22:55
Indeed
Chris Toshok
@toshok
Jul 16 2015 22:56
was confused for a second, thought I was looking at the wrong PR
@rudi-c: mind rebasing #666 off of master?
Rudi Chen
@rudi-c
Jul 16 2015 22:57
sure
Chris Toshok
@toshok
Jul 16 2015 23:01
just happened again. i was looking at 672 thinking it was 666 :)
Rudi Chen
@rudi-c
Jul 16 2015 23:02
Oh hey rebasing made the comments disappear in #666
Chris Toshok
@toshok
Jul 16 2015 23:02
yup
Rudi Chen
@rudi-c
Jul 16 2015 23:09
wat master broke
with another timeout
Chris Toshok
@toshok
Jul 16 2015 23:10
multiprocessing_test, same test that timed out for me before
I can’t remember if it was the debug build, though
I kicked it
Chris Toshok
@toshok
Jul 16 2015 23:41
k, master built after being kicked
Rudi Chen
@rudi-c
Jul 16 2015 23:41
ncie
*nice