00:08 vendethiel- joined 00:11 btyler joined, moritz_ joined 01:00 FROGGS_ joined 01:04 avuserow joined 01:23 avuserow joined 02:00 FROGGS__ joined
TimToady cut down version of t/spec/integration/99problems-51-to-60.t with a gdb stacktrace indicating GC issues: gist.github.com/anonymous/5aa8c0d7b4c99d75da35 05:19
at first I thought it might be GatherIter depending on sentinel addresses that can move, but changing them to something more constant doesn't help 05:24
not that doing address comparisons on [] isn't potentially problem, methinks, but it's not the problem in this case 05:25
or do you guys know something about [] that I don't? 05:26
(Any.new.WHICH doesn't improve things from the GC perspective, assuming WHICH is not subject to gen2 movement) 05:27
also, despite the recursive gathers, it doesn't seem to be related to the batching change that collects up to 1000 takes in eager context; changing that back to 1 still segfaults 05:31
it does seem to delay or suppress the failure if we use := instead of = when generating the trees, so eager context does bring the failure earlier somehow 05:32
nwc10 TimToady: that test fails under ASAN (unless SPESH is disabled) 06:27
(the full sized spectest) 06:28
I'm rebuilding master/master/nom with ASAN again to try the cut down version
TimToady cool, hope it finds something, though these sorts of problems are notoriously heisenbuggy 06:34
nwc10 me too, otherwise you've just found a different bug 06:36
sergot hi o/ 06:41
nwc10 \o
FROGGS morning #moarvm :o) 06:45
nwc10 TimToady: in that gist, if I just run the code in the blocks after "This is from:" ASAN built rakudo doesn't barf. 06:48
The entire spectest still barfs, with a stacktrace similar to yours
#0 0x7ffb944bac20 in scan_registers src/gc/roots.c:376
TimToady well, it's very sensitive to prior activity
that was about as far as I could cut it down here and still get it to misbehave 06:49
nwc10 is trying to read a location 0 bytes after a region allocated by allocate_frame()
FROGGS I guess that is due to our heat seeking spesh
I once disabled all spesh opts and then enabled them one by one to see what opts make it explode
at that time a combination of two were to blame, so there can be many combinations in this case :/ 06:50
nwc10 FROGGS: I tried to do something similar on PPC and went from SEGV after 8 tests (my actual bug) to instant SEGV. Is there a good way to do that disabling?
TimToady btw, my snippet still dumped with MVM_DISABLE_JIT 06:52
FROGGS nwc10: I just commented out the cases of the switch at src/spesh/optimize.c:1025 06:53
dunno if that is a good way or not
nwc10 I tried to do that.
at least, I thought that I did that
anyway, if I try it again and it fails
then I plan to use git log to work out the reverse order of which things were added to that switch 06:54
and comment out one by one
FROGGS and timotimo++ confirmed that one optimization was buggy, so he commented it out for a while (and fixed it just a week ago IIRC)
nwc10 Kaboom was on master/master/nom earlier this week 06:55
and is also in the earlist version I can get to build
so it's probably a different thing
07:02 zakharyas joined 08:06 colomon joined 08:11 colomon joined 08:55 woolfy left 11:47 nebuchad` joined 11:51 bcode joined, daxim joined 11:54 bcode joined 13:22 lastofthe joined 13:42 FROGGS[mobile] joined
lastofthe Looks like the recent malloc wrappers broke the build on ARM 32 + gcc 4.8.2, the order of the includes might need some shuffling: gist.github.com/anonymous/ce9d200f0f3686c12bff 13:57
git branch 14:03
whups
14:10 moritz_ joined
dalek arVM: 459f317 | jimmy++ | / (2 files):
fix the build on ARM 32 + gcc 4.8.2, lastofthe++
 14:11
arVM: a295bc2 | jimmy++ | 3rdparty/dynasm:
fix dynasm
 14:13
14:21 FROGGS joined, ingy joined 14:22 _sri joined
jnthn TimToady: About GatherIter, [] is used simply as a sentinel object, and it's quite clear that since [] is mutable then instances will be unique. Could be anything else wiht similar properties. I just stuck with what sorear++ picked, seeing no reason not to. :) 14:23
Also, it was used as an object pointer comparison 14:30
And those all move together while the world is stopped 14:31
14:33 17SAA04J2 joined
moritz stops the world 14:33
jnthn gets off 14:35
diakopter *cry* 15:00
tadzik quick, come back while it's still 15:02
15:15 lizmat joined 15:21 woolfy joined
TimToady will just use Mu.new then, since it's a minimal unique object, and requires no attribute storage 15:34
and is sufficient for =:= purposes 15:35
jnthn +1 15:36
TimToady [] actually has a bunch of stuff inside that is just wasted allocation
jnthn Well, it's a singleton ever so we're talking O(bytes), but yes :)
TimToady not a major thing, of course, we don't do that many gathers
there was actually a huge debate in the early days of Ada whether an empty record was good for anything :) 15:37
jnthn grumbles about his latest startup time reduction making startup longer... 15:38
TimToady discovered it was good as a peg to hang procedural code in the middle of your declaration elaboration sequence
jnthn oh, wtf...
15:42 itz_ joined
avuserow timotimo: sad news. moar segfaulted with profile-compile after some large number of hours :( 16:38
it claims it dumped a core file but I don't actually see one 16:39
oh it may not have had enough disk space to make an appropriately sized core file. 16:40
timotimo ah damn :( 16:52
but thanks for doing stuff 16:57
avuserow I added some disk space and reran it without the --optimize flag, so I'll see if that does anything 16:59
timotimo that'll cut a bunch of calls everywhere, aye 17:00
actually, without --optimize, --optimize=2 is assumed
avuserow oh 17:01
timotimo you could --optimize=off
avuserow so I should do --optimize=0?
ok
timotimo there is off and 0, i believe
if you do "off" the optimizer will be skipped entirely
with 0, i think it only does analysis
avuserow oh I should also do --gen-moar=master and --gen-nqp=master too probably?
timotimo i would do that, aye
avuserow at this point, I'm mostly curious if it's possible to complete a run and get useful data. I can tweak and re-run after that. 17:03
timotimo in theory, we could make a specialized version of the compiler that only profiles the stage "parse" and then only profiles the stage "optimize" and so on 17:04
avuserow is there any way to get the peak memory usage of a process?
17:05 zakharyas joined
timotimo sure, "time" gives you that in the "maxresidentk" slot 17:05
avuserow oh, cool. (and that appears to be /usr/bin/time, and not the shell builtin.) 17:07
timotimo ah, right, depending on the shell you may get one that's not as nice 17:11
the system time also has -v which gives you the values of different performance counters 17:12
like how many branch predictions were missed and such
avuserow yeah this is just default unconfigured bash on ubuntu server 14.04 17:14
timotimo i'm personally enjoying FISh a whole lot
avuserow I had an ISO for it laying around and it installs quickly. no other reason for using it
anyway, if the process crashed last night due to a lack of memory, it must've requested >80GB or about, since I also enabled 40GB of swap 17:16
timotimo FISh isn't a distro, it's a shell :)
avuserow right
timotimo though in the #fish channel i met a person who wanted to build a linux distro around fish
avuserow I haven't looked at fish in many years, not since I learned the basics about shells. I mostly use zsh with a nice variety of configurations 17:17
timotimo ah 17:41
i've been on zsh with lots of conf before switching to fish
recently i got the one thing back that i was lacking: alt-q 17:42
lastofthe clang throws a bunch of warnings about converting between integer types with different sign. Is "assigning to 'char *' from 'MVMuint8 *'" safe? 17:47
dalek arVM: efff2fd | (Timo Paulssen)++ | src/moar.h:
pre-declare MVM_jit_support in moar.h; fixes #135
 17:48
arVM/dynamic_gen2_tuning: 23a5c6e | TimToady++ | src/core/args.c:
Better style on arg count errors

Inverted got/expected order, also s/parameters/arguments/.		 18:05
TimToady oops, wrong branch
did we ever merge that branch? 18:07
FROGGS would be the safest to cherry-pick that commit to master 18:09
TimToady has to learn to cherry-pick, I guess
FROGGS git checkout master; git cherry-pick 23a5c6e 18:11
[Coke] note that cherry picking means merging that branch later might be rough.
FROGGS [Coke]: are you sure?
TimToady I would think that would make it easier, since it has the provenance 18:12
FROGGS I always had the impression that git knows that this commit is already there
TimToady with just applying the patch, it would have to do diffs and stuff
FROGGS for example I cherry-picked something from the labels branch, and then merged it without problems
[Coke] it doesn't know that the cherry pick is the old commit. 18:13
ok. some FUD to ignore, then. 18:14
dalek arVM: 288410e | TimToady++ | src/core/args.c:
Better style on arg count errors

Inverted got/expected order, also s/parameters/arguments/.
nwc10 did anyone do anything that was supposed to fix the ASAN failures on t/spec/integration/99problems-51-to-60.t 18:28
and the other 2 tests?
TimToady not that I know of 18:30
otoh, it seems to pass now 18:31
otgh, I was testing against dynamic_gen2_tuning, not master
and that's likely to have changed the GC characteristics 18:32
so might still be a flapper
dalek arVM/cglobal: d52b745 | (Tobias Leich)++ | / (8 files):
expose C globals like native subs
 18:34
Heuristic branch merge: pushed 21 commits to MoarVM/cglobal by FROGGS 18:35
TimToady still fails under the harness 18:38
19:07 Ven joined 19:10 zakharyas joined
nwc10 so my SEGV on PPC goes away when I comment out "case MVM_OP_getattrs_n:" in optimize_bb 19:17
investigations will continue tomorrow
but this is progress
FROGGS indeed 19:23
lastofthe nwc10: sweet. That could help my ARM twisting. 19:34
19:35 brrt joined
brrt \o 19:35
JimmyZ: what changed in the dynasm? 19:38
FROGGS brrt: it is just about your latest commit here I think: github.com/MoarVM/dynasm/commits/master 19:46
brrt oh.. i thought i had done that yet :-)
but apparantly not
19:51 itz joined
lastofthe 'though clang 3.4 will build (warningly) moar on ARMv7, unaligned reads get crashy. Here's a patch: gist.github.com/anonymous/7b1a7e81cbca6123d224 20:47
20:47 Ven joined 21:02 brrt left 21:41 itz_ joined 21:46 itz joined 21:56 itz_ joined 22:49 lastofthe joined
lastofthe clang on arm troubles filed with patch as #137. 22:57
avuserow A friend wants to know if there is a SuchVM since there is a MoarVM. 23:17
avuserow .oO( SuchVM. Very Execute. Wow. )
lastofthe wonders where Configure.pl might put the compiler's -W flags... 23:31

Powered by Cro and the Raku® Programming Language.

Please report any issues / comments / feature requests as an issue on App::Raku::Log.

Thank you!