| IRC logs at
Set by AlexDaniel on 12 June 2018.
00:37 AlexDani` joined 00:39 AlexDaniel left 01:59 Util left 02:22 Util joined
discord6 <sjn 🇳🇴> \o 08:44
lizmat sjn o/ 08:49
discord6 <sjn 🇳🇴> Coveaster hacking! \o/ 08:52
09:12 Altai-man_ joined 09:22 sena_kun joined 09:23 Altai-man_ left 09:37 angelds joined 11:20 angelds left 11:21 Altai-man_ joined 11:23 sena_kun left 11:35 AlexDani` is now known as AlexDaniel, AlexDaniel left, AlexDaniel joined 11:38 MasterDuke left 11:41 patrickb joined 11:48 MasterDuke joined 12:40 samcv left 13:22 sena_kun joined 13:24 Altai-man_ left, zakharyas joined 13:43 samcv joined 13:51 samcv left 13:53 samcv joined 14:20 zakharyas left 14:56 samcv left, samcv joined 15:21 Altai-man_ joined 15:23 sena_kun left 15:24 colomon left 15:26 colomon joined 15:36 colomon left 15:38 colomon_ joined 16:25 zakharyas joined
timotimo i think the profile dumper should probably stop when the recursion depth has exceeded like 20k 16:28
at that point you're very likely in a situation that isn't going well
MasterDuke timotimo: true. but i also feel like in some/most of those cases the actual code that the profiler is collecting info on is not great 16:42
timotimo it's more likely that it's still that nasty bug that causes loops to turn into infinite-depth recursion in the eyes of the profiler 16:49
MasterDuke is it strictly a bug in the profiler? 16:57
or some badly generated/optimized code that the profiler just doesn't handle well? 17:00
timotimo it's a bug that i haven't properly figured out, but if i have the right idea, it's that when generating the enter/exit instructions in the code, sometimes there's a path through the routine that skips the exit instruction on the way out 17:04
MasterDuke that sounds very plausible. might explain why when trying to compile big things like the rakudo build i never see it ending profiling 17:05
17:22 sena_kun joined 17:23 Altai-man_ left 18:30 vesper11 joined
timotimo who is bored enough to try putting liborc into moarvm? 18:30
it uses meson as its build tool, so have fun with that i guess
18:30 vesper11 left
timotimo but it offers an assembly-like language that it can compile to vectorized code targeted at the compile-time or run-time selected architecture, or you can put the programs together with regular C function calls 18:31
it'd be nice to have liborc as optional if it is to be included 18:32
and then having the nqp level check if liborc is there or not and have the alternative implemented simply in nqp 18:33
OTOH, composing a QAST in raku code and compiling it, then running it could already be a nice win for very big arrays being hyperopped 18:36
18:50 vesper11 joined 19:21 Altai-man_ joined 19:23 sena_kun left
timotimo oh my, a spesh mystery 19:29
MasterDuke do you have any dogs and/or meddling kids around to help solve it? 19:40
timotimo i've been barking up the wrong tree for a little there ... 19:51
nwc10 MasterDuke++ 19:52
timotimo there we go. the self for this method wasn't restricted to :D, that's why the getattr couldn't be speshed 19:53
and it's a certain optimization, possibly because it's only got osr hits 19:54
MasterDuke what are you optimizing? 19:55
timotimo was looking a bit into our hyperop implementation 19:56
MasterDuke cool 19:58
timotimo if there were a nice little spot where i could say "we have an int array on both sides and a simple operator from the setting, so let's use a more optimized approach here" 19:59
and then i stumbled upon a getattr_o that was looking like it should have been optimized 20:00
and i didn't see why until i put in more debug info
now i'm storing an extra integer in spesh comment annotations that stores the order in which they were added 20:01
MasterDuke what does that tell you? 20:03
timotimo i wanted to make sure it was doing the one thing that i saw before the other thing that i saw 20:05
because i was being misled by buggy debug output :)
i'll post an example 20:07
it also gives a hint for stuff that has been deleted by spesh, since the comment numbers will have a gap 20:09
theoretically i could save them up and display them after the graph with what they used to attach to 20:10
MasterDuke huh, interesting 20:19
20:48 zakharyas1 joined 20:50 zakharyas left 21:22 sena_kun joined 21:23 Altai-man_ left 21:30 zakharyas1 left 21:57 patrickb left 23:21 Altai-man_ joined 23:23 sena_kun left