On 23.9.2014 2:10, Dmitry Marakasov wrote:
* Dmitry Marakasov ([email protected]) wrote:
I've been getting some failure mails from pkg building cluster related
to clang crashes:
http://beefy2.isc.freebsd.org/bulk/head-amd64-default/2014-02-28_01h43m56s/logs/arx-libertatis-1.0.3_2.log
http://beefy1.isc.freebsd.org/bulk/head-i386-default/2014-02-21_03h01m36s/logs/supertuxkart-0.8.1.log
At first I thought they'll go away with clang 3.4, but recently I've
added HEAD jail to my tinderbox and reproduced these. Also a reason
of crashes become apparent: c++ eats all available memoty (15G RSS
in my case).
I though of investigating these further, but wanted to check if it's
already known first.
There are a few known OOM bugs in 3.4, but from the logs it is not
immediately apparent if you are hitting those, or if they are new bugs.
Note that it affects both 3.3 and 3.4. I think it was triggered by
something unrelated to clang update, maybe libc++ update.
To be able to figure that out, we need the files mentioned in clang's
diagnostic messages, e.g. for arx-libertatis:
/tmp/Input-XkKTNq.cpp
/tmp/Input-XkKTNq.sh
Here's one from arx-libertatis:
http://people.freebsd.org/~amdmi3/clang-eats-mem-bug.tar.bz2
The bug is reproducible on my 10.x with both system clang 3.3 (after
removing -vectorize-loops -vectorize-slp options) and with clang 3.4
from ports.
I'll just remind that this is still an issue, and it's getting into
10.1. Could we maybe disable compiler features with lead to these
crashes?
This seems to be same issue as in
http://llvm.org/bugs/show_bug.cgi?id=20893 for which there is patch
review going on
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20140922/236415.html.
Your test case is reproducible with the trunk of llvm/clang and the
patch in review resolves it. Other workaround is to disable generation
of debug information by removing -g flag.
Pasi
_______________________________________________
[email protected] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-toolchain
To unsubscribe, send any mail to "[email protected]"
