Welcome to the main channel on the development of MoarVM, a virtual machine for NQP and Rakudo (moarvm.org). This channel is being logged for historical purposes.
Set by lizmat on 24 May 2021.
00:02 reportable6 left 00:06 reportable6 joined 00:53 shareable6 joined, evalable6 joined 00:54 bisectable6 joined, benchable6 joined 00:56 tellable6 joined 01:12 tbrowder left 01:15 tbrowder joined 01:46 frost joined 01:53 greppable6 joined 01:54 quotable6 joined 02:56 unicodable6 joined 04:56 nativecallable6 joined
nine m: my $done = 0; class Foo { submethod DESTROY { $done++ } }; loop { VM.request-garbage-collection; Foo.new; last if $done; } 05:52
Hah....seems to hang on master as well
camelia (timeout)
05:55 statisfiable6 joined 06:02 reportable6 left 06:05 coleman-x joined 06:07 colemanx left 07:05 reportable6 joined 08:13 frost left, frost joined 08:14 frost left, frost joined
Nicholas good *, #moarvm 08:36
MasterDuke i'm trying out controlflag (github.com/IntelLabs/control-flag), intel's new static analysis tool. training on the sample data they provide took 500s. now it's scanning moarvm while pegging all my cores. top says 435:00.84 for time so far... 08:42
it's only scanned 94 files so far and i just realized i pointed it at the repo base directory, not ./src. i see it going into the .git directory, stopping now... 08:47
i think maybe i'll leave this tool for someone with a beefier machine than i have 08:48
jnthnwrthngtn moarning o/ 09:15
Nicholas \o
Geth MoarVM: MasterDuke17++ created pull request #1574:
Improve the jit implementation of nqp::abs_i
09:41 squashable6 left 09:42 squashable6 joined
jnthnwrthngtn Goodness it's windy here today... 09:46
lizmat as it is here, fwiw :-)
jnthnwrthngtn Think I have a fix for the regression in SupplyTimeWindow
A building across the street has been having a floor added and thus a new roof. The sheeting covering the roof to protect it from the rain just blew off. 09:47
At least it's only windy, not raining...
lizmat plenty of trees didn't make it through the night in NL 09:49
still have quite some leafs on them, made them less aerodynamic 09:50
jnthnwrthngtn yay, spectest is clean 09:53
lizmat and more likely to succumb
jnthnwrthngtn oops, I'm meant to be at a planning session for our hall...bbiab
11:21 linkable6 left, evalable6 left 11:24 linkable6 joined 12:03 reportable6 left 12:22 evalable6 joined 12:51 linkable6 left 12:53 linkable6 joined
sena_kun jnthnwrthngtn++ # fixing things 12:55
13:06 reportable6 joined 14:00 Kaiepi left 14:11 TempIRCLogger left 14:16 lizmat left 14:54 Kaiepi joined
[Coke] HECKIN WIMDY 15:21
Nicholas it blew one of our plant tubs over. Which turns out to need watering. But still 15:26
this has not happened before
(fairly big tub, fairly small plants. Does not look like an obvious sail)
jnthnwrthngtn Our terrace gets quite windy anyway, so when I shopped for plant pots I went for heavy ones. 15:31
nine I think I know why those destructors don't get called in the version with $*VM.request-garbage-collection. 15:32
request-garbage-collection does nqp::force_gc, which enters the garbage collector immediately. The GC collects the garbage and creates a list of finalizers to call. Since this list is not empty, it sets up a special return handler for the current call frame which should call the finalize_handler_caller 15:34
The special_return handler is part of the frame's ->extra 15:35
After the GC run, request-garbage-collection returns to its caller. 15:36
The return_o op calls MVM_frame_try_return which in turn calls remove_one_frame. 15:37
Now look at what remove_one_frame does as its very first action: github.com/MoarVM/MoarVM/blob/mast...ame.c#L885
The comment says it all: /* Clear up any extra frame data. */
But the extra frame data also contains the special_return handler which is supposed to call the finalize_handler_caller! 15:38
So when are special_return handlers ever called you may ask. Well, later in remove_one_frame: github.com/MoarVM/MoarVM/blob/mast...ame.c#L952
But we're running the special_return handler in the caller frame's extra data! 15:39
So we only ever run finalizers if the frame that happens to trigger a GC run makes any call afterwards. If it doesn't make a call but returns, we do not run the finalizers. 15:40
jnthnwrthngtn I really wish we could find a different way to do this. I want to kill off special return handlers as we have them today in favor of a new kind of call stack record, and we'd thus run them on unwinding and have them "for free" rather than as an explicit check for them at frame exit 15:42
Finalization is the only case I know of where we can't do that 15:43
nine And it's not just that we don't run those finalizers. We also keep an ever growing list of objects to finalize
jnthnwrthngtn Because it is also the only case where we don't set up special return things ahead of the invoke 15:44
But rather want to place it down the stack
MasterDuke and it's not as simple as running what's in the body of that `if (caller->extra) {` before what's currently the body of `if (returner->extra) {`?
nine Yeah as it is now it sucks. As it is we don't run them unless the frame makes some call between GC and exiting. If we put it into the caller's extra, we may never run them if we don't return to the caller
jnthnwrthngtn So let's not invest time fixing the current way, because I'm trying to replace it :) 15:45
I wonder if we can do the same technique as we have for lazy deopt :)
Except it kinda needs to compose with that so we need to be careful
nine A workaround for now could be to simply add a call to request-garbage-collection after nqp::force_gc. May just be tricky to ensure that this all doesn't get inlined away 15:46
jnthnwrthngtn But it twiddles the kind of a record on the stack so that when we return into it, it doesn't appear as a frame, but instead as something else, which we then figure out is a frame that needs deopting.
(there's an original kind for this) 15:47
Since we're in GC we know we're not running code so it's safe to do this manipulation
The bug not to repeat is to do it on the topmost frame, but instead its caller
So that the unwind logic will see what's going on 15:48
nine jnthnwrthngtn: why is this tied with returning in the first place?
jnthnwrthngtn nine: Because it's a safe point to invoke something else 15:49
And points where GC may occur are in general not 15:50
And we've ended up with a dependency on doing finalization on the same thread as the objects were allocated on.
Also because we already have some mechanism for doing something differnet upon return (and in fact rely heavily on this) 15:52
nine If we tie it to returning to the caller what happens if the GC is triggered from a loop we don't return from?
May osrpoints help?
jnthnwrthngtn Tough
No, because they're gone in specialized/JITted code. 15:53
From the perspective of Inline::Perl5, what matters? That the same thread releases the objects, or that there is no concurrency between DESTROY and the thread?
nine That the same thread releases the objects
jnthnwrthngtn OK
That means we can't have a fianlizer thread and refuse to let any of the others continue until it has done its work. 15:54
Though that'd be very fraught too
nine Which precludes concurrency :)
jnthnwrthngtn Well yes, I was pondering if we could have our cake and eat it in that sense
(A separate thread so we know it's at a safe point for invoking something, but without concurrency with the other threads) 15:55
I mean, "no concurrency" is a weaker condition than "on the same thread" :) 15:57
I guess we could try something like: tie such a check to the osrpoint in unoptimized code, and in optimized code, if we have a osrpoint, and the loop that it appears in will never call anything, then rewrite it to something else (a "deopt if we have to run finalizers" or some such op) 16:01
That only needs loop detection and abstract interpretation, how hard can it be... :)
(Though we can cheat and look for the backend after the osrpoint) 16:02
language class, bbl
Nicholas nine: IIRC mod_perl was crazy enough to re-map which perl interpreter structure was associated with which OS thread. So in theory that can be done 16:24
but life is probably easier it one keeps OS thread and perl interpreter structure with the same 1:1 mapping. Because I don't know if some XS code assumes this. And never really had to run under whichever mod_perl variant did the crazy stuff 16:25
nine Nicholas: I'm not sure I am crazy enough... I mean, I'm plenty crazy, no doubt about that. But even I may have my limits :) 16:36
The only realistic alternative I see is give wrapped Perl objects a reference to their interpreter instead of tying the latter to the Raku thread. But so far I've managed to keep performance of single threaded uses untouched and I'm not sure I could keep this up. 16:38
About finalizers: I dare say it's quite common for a routine to do something that may trigger a GC run but not call anything before returning. At least compared to frames that may trigger GC but never return (or not for a very long time). 16:52
So its safer to tie finalizers to the caller or to returning to the caller, rather than to the current frame or returning into the current frame. 16:53
Worst case for that "endless" loop is running finalizers at program exit. Still bad (as it may lead to unbounded growth), but at least somewhat deterministic. 16:54
If we want to improve that via osrpoints, there's no use in checking for whether the loop calls something (as we'd run finalizers when exiting the frame the loop's in). 16:56
I wonder what makes returning a safe point to invoke something else? 16:58
timo hm, i wonder if we'll ever generate code so tight that we can mark a thread as unable to participate in GC while it does its whatever. like heavy number crunching? 16:59
it could probably only work with native locals in that case
nine timo: goto is a GC_SYNC_POINT
Hard to create loops without goto :)
timo make a goto2 then 17:00
like, with sufficient scalar replacement, a mandelbrot renderer could do all its calculations without touching gc-managed objects. it'd just need to put its results into like an array or something, which we don't have a scalar replacement story for yet. at the very least not for anything really big. 17:03
funnel the results out through NativeRef :D 17:04
i guess it's just a little bit of a pipe dream ;) 17:05
nine You think that now 17:19
timo for anything that wants to be interactive, say a video game perhaps, gc pauses will remain a problem in the future 17:24
nine Well, yes and no. Nasal (scripting language used in the FlightGear simulator) also has a stop the world GC: wiki.flightgear.org/How_the_Nasal_GC_works 17:39
timo stop the world is fine if you can get pauses to be at most n ms long, like with incremental GC 17:45
17:47 TempIRCLogger joined 17:53 lizmat joined 18:02 reportable6 left 18:05 reportable6 joined
lizmat waves from Cologne 18:16
gfldex lizmat: don't let go of your hat!
It's very windy in germany right now.
timo kölöln 18:30
nine: i didn't actually read the complete page. were you pointing something particular out? they're running some scripts in separate processes (i guess you could have multiple nasals in the same process with different threads and split pools up and such? maybe they even do that already) and use IPC 18:32
nine timo: no nothing in particular. Only provided the link for the curious. Nasal is definitely running in the main process. Easy to tell since Flightgear just doesn't spawn any other processes at all. 18:41
timo oh, OK 18:47
18:55 TempIRCLogger left
lizmat goes back home 18:56
18:56 lizmat left
[Coke] ... wait, you guys get to leave your house?? 19:42
gfldex hands [Coke] a ⁇ 19:46
[Coke] Спасибо 19:51
Nicholas For the avoidance of doubt: perldoc.perl.org/5.35.5/perldelta -- for my ($left, $right, $gripping) (@moties) { ... } 20:02
timo i don't even know what moties are 20:21
[Coke] to me, moties are spacebattles-factions-database.fan...iki/Moties 20:24
... which now that I reread the other vars, is actually those moties too 20:25
Google "mote in god's eye" for the first book, plenty of pics & plots. 20:37
20:44 lizmat joined, TempIRCLogger joined
japhb [Coke]: This line caught my eye further down: "Unrelated changes accidentally broke the build for the NetWare port in September 2009, and in 12 years no-one has reported this." 20:56
I didn't even know NetWare had made it into this millenium 20:57
Oh, I want to second timo's comment: For games and animation, if I had bounded GC pauses, where the bound was (in the best case, significantly) less than half a frame time, I'd be fine running it every frame. It is *far* more annoying to have random stop-the-world hiccups than to have to mildly reduce world complexity to keep frame rate up. 21:04
(Or do some extra optimization, of course) 21:07
timo japhb: did you have a look how the terminal drawing library performs with new-disp? 21:12
Terminal::Print i think was the name? 21:14
japhb timo: Oh, no I haven't. I'll build a head and a 2021.09 and see what's the difference 21:33
Oh, is there a pending bump? ISTR jnthnwrthngtn and lizmat talking about it. 21:34
japhb builds 2021.09 first, since the answer to the bump question doesn't matter for that one. 21:35
lizmat jnthnwrthngtn appeared to not want a bump
otherwise I would have done one already
japhb lizmat: Ah, yeah, that fits my fuzzy memory
OK, well it turns out I'd saved a 2021.09 build from right after the previous release, so I guess it's back to doing Rakudo HEAD and see what I get 21:36
21:37 vrurg_ joined, vrurg left
jnthnwrthngtn lizmat: I'm a bit confused; I expected it to happen as part of the blin run being performed, but actually we have blin results but no bump... 21:38
lizmat well... colour me confused as well then :-)
japhb feels less stupid now
jnthnwrthngtn Anyway, feel free, I don't plan anything more for master in any of the repos this side of the release 21:40
japhb BTW, got some progress from the CBOR standards folks: It looks like they're convinced that supporting Capture in CBOR natively is a good idea, so now it's just quibbling over some details.
(I mention that here, because one of the use cases I discussed was high-volume tracing and debugging streams) 21:41
jnthnwrthngtn Ah yes, need to do the Log::Timeline work on that
japhb :+1: 21:42
jnthnwrthngtn Well, and the Comma side of it. Should get a few of the nice in-progress things finished up there first, though :) 21:45
sena_kun if there was no bump after the morning fix, then it is necessary. :) 22:10
japhb Well that doesn't bode well: MoarVM oops: MVM_str_hash_fetch_nocheck called with a stale hashtable pointer 22:39
(That's from really deep in the Terminal::Print library, though -- no golf yet)
Yep, it's reliable on my multi-threaded case. Lemme try single-threaded. 22:40
Single threaded worked. Here's the result for current HEAD rakudo: 22:45
Min: 79.2 ms (12.6 fps) - Ave: 208.8 ms (4.8 fps) - Max: 403.7 ms (2.5 fps)
50%: 206.5 ms - 75%: 255.1 ms - 90%: 281.1 ms - 95%: 292.4 ms - 99%: 375.9 ms
(That's statistics on frame times of the most complex Terminal::Print animation test I have) 22:46
OK, trying again on 2021.09
Min: 157.2 ms (6.4 fps) - Ave: 330.4 ms (3.0 fps) - Max: 562.8 ms (1.8 fps) 22:50
50%: 327.0 ms - 75%: 399.6 ms - 90%: 425.1 ms - 95%: 441.9 ms - 99%: 531.8 ms
So new-disp is significantly faster (lower frame times, higher fps) for that test.
Let's see how mandelbrot-pixels does 22:53
2021.09: 16 zooms of 62460 pixels each in 7.325 seconds = 136437 pixels/second 22:54
HEAD is weirdly variable. In three runs I got: 22:55
16 zooms of 62460 pixels each in 5.749 seconds = 173846 pixels/second
16 zooms of 62460 pixels each in 6.814 seconds = 146660 pixels/second
16 zooms of 62460 pixels each in 6.744 seconds = 148174 pixels/second 22:56
(Multiple runs of 2021.09 were less than half a percent different.)
So summarizing: Can't do multithreaded animation test because of MoarVM oops; HEAD significantly faster for `raku -Ilib examples/attacks.p6 --show-fps --bench --slow-mo=10` in a maximized terminal window, HEAD faster but unstable performance wise on mandelbrot-pixels test. 22:59
MasterDuke i think that oops is probably because of concurrent hash modification, which isn't allowed 23:03
japhb MasterDuke: Hmmm, that may be concurrent updates to the color mapping cache. I may be able to make a per-thread cache for that, at the cost of some memory efficiency. (Or for low bit depths, just cache all the possible cases.) 23:19
23:35 linkable6 left, evalable6 left 23:36 linkable6 joined