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.
00:03 reportable6 left 00:05 reportable6 joined 06:02 reportable6 left 06:05 reportable6 joined
Nicholas good *, #moarvm 07:00
Nicholas follows the URL from last night.
Today I learned that "session mead" was a thing.
nine With my deserialization fixes t/05-autocomplete.t ran successfully in a loop for 10 hours or so. Then it crashed with "corrupted double-linked list" message by malloc when resizing an array during deserialization. 07:32
Nicholas Gah! And of course you have no idea whether you tickled a different bug. 07:33
nine It's random memory corruption. Could be the issue I was chasing, but also something else. No way to determine that from just a core dump. 07:35
Nicholas you need (at least) ASAN, if not get lucky with rr. Aaargh, Aaargh
nine The 10 hours are a good sign anyway. After all, I do have a reasonable explanation for why my changes fix bugs. They just don't fix all of them :) 07:39
At least the weather is suitable today for running stuff in loops during the day :) 07:43
So I ran it with asan, but of course it complains about leaks. Add --full-cleanup and bang! "Collectable 0x623000000130 in a gen2 freelist accessed" every single time. 07:50
Nicholas oh, that 07:51
yes, I've seen that.
nine But use of gentle force allowed for me to reproduce it in gdb and it's just the missing worker thread cleanup in --full-cleanup that's causing this. Maybe I should continue my work on that some day
jnthnwrthngtn moarning o/ 09:07
Nicholas \o
jnthnwrthngtn Time to find out what happens if I switch the QAST compiler to compile call and callmethod to use new-disp.. 09:17
Nicholas make sure you have good backups!
candles, matches and a flashlight 09:18
.oO( generator for the beer fridge )
I realized that I don't have to swich both method calls and normal calls at the same time, so currently going for method calls. 10:09
Nicholas jnthnwrthngtn: or simply, "a deep enough wine cellar such that it maintains a good temperature with zero carbon footprint" 10:11
jnthnwrthngtn Argh. So we get some of the way then...run directly into the bootstrap 10:20
Nicholas does that make it lunchtime? 10:21
jnthnwrthngtn 'cus of course there isn't one NQPRoutine to consider, there's two
Nicholas and this won't get easie runless you actually do both at once? 10:23
10:30 AlexDaniel joined
jnthnwrthngtn Well, there's an easy hack while I ponder if there's a better solution :) 10:34
Then there was an ordering issue. And now I'm running into not having ported multiple dispatch yet :) 10:39
lizmat hopes jnthnwrthngtn will be able to write a cool blog post as a cooling down exercise :-) 10:47
jnthnwrthngtn Hm, this isn't quite so painful as feared. With QAST -> MAST compilation compiling callmethod nodes into the dispatch op, NQP builds and fails only 13 test files 11:03
Well, I don't know how painful fixing those will be :)
Geth isn't reporting still, but I did a temporary thing that lets me get the single dispatch ported and working and then do the multiple dispatch after. 11:04
lizmat I'm afraid recent RSC events caused Geth to become unresponsive :-( 11:09
jnthnwrthngtn Ah, OK
.oO( Raku Secret Commits )
lizmat I'm in the middle of preparing my presentation at esLibre tomorrow, so have no time to look at this now
jnthnwrthngtn It's fine, I'll survive :)
I'm meant to work on something else this afternoon anyway. 11:11
Nicholas that's OK too - afternoons don't actually exist
nine Because....noon doesn't exist, right? 11:12
Nicholas I'm not sure quite why. Lunch still exists.
but maybe it's just one sort of "all day breakfast"
nine One can have lunch in the morning
jnthnwrthngtn oh, right, lunch
I didn't push the commit that switches over callmethod; will try and debug some of the test issues first 11:14
Maybe more of that later today. Dunno. We'll see how quickly I tire of this afternoons's Go + strangely documented C API advernture. :)
*adventure, even
Nicholas Given that choice, it sounds very tempting to "out to lunch" and stay there. Pesky COVID, spoiling these plans... 11:15
12:02 reportable6 left 12:03 reportable6 joined 14:17 dogbert17 left
nine I don't know what it is about this test but it seems to inevitably end in some crash when run often enough. Now up: segfault in spesh stats add_type_at_offset 14:33
Cought in rr this time :) 14:34
Certainly doesn't look right. The computer may be excused for getting confused by this: $4 = {bytecode_offset = 542, num_types = 1426, types = 0x4, invokes = 0xffff00020000021e, num_invokes = 542, num_type_tuples = 1426, type_tuples = 0x10, plugin_guards = 0xffff000000000592, num_plugin_guards = 1904} 14:35
It looks so weird because it's actually a static frame's handler: {start_offset = 542, end_offset = 1426, category_mask = 4, action = 0, block_reg = 2, goto_offset = 0, label_reg = 2, inlinee = 0} 14:39
15:05 frost left
jnthnwrthngtn Well, that's pretty broken. 15:19
nine Considering that MVMSpeshStatsByOffset are malloced, this means that it must have been freed prematurely, which would happen when the MVMStaticFrameSpesh holding it gets collected. Trying to verify that now 15:26
Alternatively it could also be that handler that got freed prematurely and its memory re-used for the by_offset. 15:28
If my break point on MVM_free with the MVMSpeshStatsByOffset's address doesn't hit, that's probably the case 15:29
jnthnwrthngtn Turns out a lot of the failing NQP tests were due to testing things that are changing or relying on things that are changing incidentally; after fixing those up, there's only 4 test files pointing out almost certain problems. 16:57
And 2 are the same exception 16:58
lizmat wow, that's good progress! 16:59
jnthnwrthngtn One MoarVM fix later, down to 3. One of them is because I didn't reinstate the method not found error reporting callback in the new dispatcher yet. 17:31
And the other two are a compiler blow-up in termish, which is weird in so far as we must go through that tens of thousands of times without trouble elsewhere. 17:32
Think that's for after dinner. 17:34
nine Tens of thousands of times? Are you running with spesh enabled? 17:38
So...the break point on MVM_free did not fire 17:39
lizmat so what is the "<" twigil ? 17:42
docs says: "<Index into match object (not really a variable)"
but is no further docs 17:43
(yes, working on a presentation on sigils and twigils) 17:44
nine But....the point where we overwrite the MVMSpeshStatsByOffset is immediately after allocating the handler array. So it can't be a freed handler 17:49
jnthnwrthngtn nine: No, I just mean that given we compile the entire of NQP and all but these 2 tests, we must go thorugh the termish rule a lot of times without trouble, so why does it break this time? 17:51
(Not actually asking you. :))
Or rather, in exactly 2 cases
lizmat: It's not really a twigil, but I think it's about $<foo> 17:52
lizmat aha... ok
jnthnwrthngtn food, bbl 17:53
nine Wait a minute...this already doesn't look exactly sane: (rr) p *tss 17:58
$20 = {arg_types = 0x0, hits = 37, osr_hits = 0, by_offset = 0x47e8d70, num_by_offset = 67108865, max_depth = 1}
Unless 67 million entries in that array are to be expected. Which I doubt. If nothing else, from the fact alone that we grow this array one entry at a time, which we ususally don't do if we expect it to grow a lot 17:59
Nicholas I'm not sure how to be a more helpful rubber duck 18:00
sure, quack!
but I don't know what useful question to add
I guess I can ask if your beer fridge is well stocked. 18:01
Or are they mead fridges these days? :-) 18:02
oh yes. sherry. Let me test that...
18:02 reportable6 left
nine Ooooh...the plot thickens. That fishy tss thing, we get it via tss = simf->type_idx >= 0 ? &(simf->ss->by_callsite[simf->callsite_idx].by_type[simf->type_idx]) : NULL; 18:05
18:05 reportable6 joined
nine Now simf->type_idx is 8. But! simf->ss->by_callsite[simf->callsite_idx].num_by_type is just 4! So it's an out of bounds array access. 18:06
The bad news is that this means we have some logic error and I'm not terribly familiar with spesh's stack simulation code. 18:07
18:21 dogbert17 joined
dogbert17 nine: seems like you fell into a rabbit hole :) 18:22
nine Noooo...that never happens! Would be a complete first
dogbert17 so have many bugs have you fixed, one, two or more 18:23
MasterDuke looks like it's been a busy day 18:40
github.com/faster-cpython/ideas/is...dated-desc has some interesting stuff we might be able to steal/be inspired by 19:35
22:24 m6502 joined 23:16 m6502 left