| IRC logs at
Set by AlexDaniel on 12 June 2018.
03:00 sivoais_ left, sivoais joined 06:59 patrickb joined
nwc10 good *, #moarvm 07:10
jnthn: OK, so, I had a bit of inspiration on how to attack the problem 07:11
turns out that spesh assumes that hash entries don't move once they exist
specifically sp_gethashentryvalue
I'm not sure that I like this assumption. I think (not yet tested) that it can remain valid by introducing a level of indirection. But, yes, lots of 8 bytes and cache misses 07:12
but 07:13
I think also that it's possible to product 2 branches - one where hash entries move and that spesh optimisation is disabled, and one with indirection where it is not
and (if both get to the point of passing)
nine If in doubt, remove that spesh optimization. It's just saving one hash lookup. If hashes become faster, the opt wouldn't do so much good anymore anyway :) 07:29
nwc10 well, I don't yet know what else we would hit 07:32
I've found this by running under ASAN, and malloc/copy/free the hash contents on each insert or delete
nine I don't think anyone else was audacious enough to rely on this... 07:39
nwc10 we will see (at least as far as spectest) 07:40
but not yet, as I have a bit of $ork to get done
07:40 MasterDuke joined
MasterDuke nine: i thought the opt saved two hash look ups and a lock? 07:41
nine MasterDuke: only part of that relies on the hash entry not moving IIRC 07:42
MasterDuke oh, yeah, that makes sense 07:43
nwc10 OK, interesting, it's: 07:50
MVMHashEntry *entry = MVM_str_hash_fetch_nt(tc, hashtable, sym_facts->value.s);
if sym_facts isn't visible to NQP code, then *it* can become a hash-with-indirection
and the rest not
but let's leave that as a future idea
and yes, that might mean in the future that MVMHash needs to fork into a second variant with the indirection 07:51
if we do want these sorts of hashes visible to NQP
07:51 sena_kun joined
nwc10 sena_kun: I might have found it (my problem). See recent scrollback. 07:52
but I have to babysit the work $thing for a bit
so can't hack on this
or the linux isntall on the laptop to my right
this tank of a laptop that has a BIOS sufficiently old that it won't USB boot :-( 07:53
its future role is "USB disk caddy" :-) 07:54
(well, mobile device photo backup target)
oh, I'm misreading that line 07:59
it's *hashtable* that matters
and that is visible to the world
anyway, the $ork thing has a new bit of logging, and, interesting.... 08:00
sena_kun nwc10, sounds interesting, good luck with sorting it out. :)
tellable6 2020-07-16T21:24:52Z #raku-dev <patrickb> sena_kun: I'm totally stumped on this one! This builds fine locally and all the tests are green. No idea why nuking should be necessary, but did you try that?
nwc10 also, I realised that there's a potentially (somewhat nutty) optimsation possible with going to a *single* memory block for a hash (including the control structure in that block) 08:04
hash copy is
1) malloc
2) memcpy
3) fix up the write barriers
08:24 vrurg left 08:25 vrurg joined 08:29 vrurg left
nwc10 bother. that's not it 08:40
09:12 Altai-man_ joined 09:14 sena_kun left
jnthn morning o/ 09:30
nwc10 \o 09:31
09:48 vrurg joined 09:49 vrurg left, vrurg joined 11:10 dogbert17 joined 11:13 sena_kun joined 11:14 Altai-man_ left 11:44 vrurg left 11:49 vrurg joined 11:54 vrurg left
nwc10 dear @all 11:56
so my error message is this:
Method 'hash' not found for invocant of class 'Capture'
IIRC I know which rakudo source file that gets reported from 11:57
but underyling this - what is actually going wrong? Where in the C code should be looking?
as in "I see that we throw this from method_not_found_error in src/Perl6/bootstrap.c/BOOTSTRAP.nqp" but how did we get there? 11:59
jnthn Full stack trace? 12:02
Exception handlers run on the top of the stack before unwinding, so in principle we can see how we got there
Methods resolution currently fist looks in the method cache hanging off the STable
*first 12:03
Which is a hash
nwc10 12:04
"is a hash" ought to be the cause here 12:05
but it's now built with some agressive reallcation
(at least it should be, and it found the thing in spesh in the nqp build)
but that backtraces is with MVM_SPESH_DISABLE
this will all be down somewhere in MVM_6model_find_method() ? 12:13
jnthn Yes 12:19
And the hash in question is the method cache one hanging off STable 12:20
nwc10 Which gets deserialised, it seems, or at least, can do
but I was curious which bit of C code I needed to hack in some extra "logging"
but "it's not spesh, it's not the subtle one of taking the address of the HashEntry struct (because ASAN triggers on that), and it gets 100% through NQP" 12:22
and it's not (other) undefined behaviour
jnthn I'd probably instrument find_method
nwc10 (oh wait, or is it. this hasn't yet been attacked with valgrind since I added the agressive malloc/copy/free)
jnthn (the one you mentioned)
nwc10 OK, yes, that looks like the candidate
that or "attack the weeds in the garden"
jnthn And see what lookups in the cache fail/succeed 12:23
12:30 vrurg joined
jnthn Friday afternnon really isn't the time to dig back into the dispatcher stuff...guess I'll give it a couple of days next week. 12:46
lizmat use nqp; my $l := nqp::list_i(42); dd nqp::existspos($l,1) 12:50
evalable6 0
lizmat use nqp; my $l := nqp::list_i(42); dd nqp::existspos($l,0)
m: use nqp; my $l := nqp::list_i(42); dd nqp::existspos($l,0)
camelia MVMArray: atpos expected int register
in block <unit> at <tmp> line 1
lizmat weird, huh ?
jnthn I don't think existspos actually makes sense on a non-object list 12:51
nwc10 jnthn: were you planning to POETS, or just to find something shorter to do? 12:52
lizmat jnthn: you're right, but shouldn't it explode as well if the element does not exist ?
jnthn lizmat: existspos is sugar for elems + isnull(atpos(...)) 12:53
And elems comes first
lizmat yeah, figured it'd be something like that 12:54
jnthn nwc10: I'll do some more RakuAST bits, 'cus that's mostly swapped in still :)
nwc10 ah righto 12:55
jnthn Also figure I should get some work in on things, 'cus I've got 2 weeks afk vacation coming up :) 12:56
nwc10 cool. 12:57
jnthn (From Thursday)
nwc10 you plan to "head for the hills" before the release :-) 12:58
13:12 Altai-man_ joined 13:15 sena_kun left
timotimo jnthn: looking forward to your next dispatcher hacking :) 13:19
.oO( or is it "run to the hills"? :) )
grr, whichever end I pick to unravel next, I find it has a yak attached... 13:38
timotimo yaks love to do tug-of-war, i assume that's what's going on here 13:39
moritz drink a yakuccino, maybe it's better afterwards :D 13:40
jnthn Turns out I really need to figure out the whole implicit topic thing
I actually wanted to do exception handlers next but those end up implying the same thing :) 13:41
nwc10 does the beer fridge have a yak attached? 13:47
or hiding inside?
timotimo ah, this is RakuAST work 13:48
jnthn I'm quite sure the beer fridge is yak free 13:49
timotimo: yes 13:50
Well, by this point I have a newly written failing test and a cup of tea...
nwc10 "I'm quite sure the beer fridge is yak free" -- that sounds like an interesting challenge for April 1st. Would probbly need an inside collaborator
timotimo jnthn: is there any specific reason why we can't move getcode + takeclosure closer to before we invoke/use the thing for the first time? 13:58
like, i would assume we can skip doing that for if/elsif/else branches where we only invoke one of the n 14:00
jnthn Well maybe but then what if it's in a loop? 14:01
timotimo right, we'll have to move it before the first branch that is relevant to the usage site i guess 14:04
the latest block that dominates it perhaps 14:05
that doesn't seem right
the latest block that dominates both it and the exit block?
there's certainly prior art for this simple question :) 14:07
put it right in front of the thing, then do loop-invariant-code-motion or whatever 14:08
14:18 AlexDaniel left 14:21 vrurg left 14:24 vrurg joined
nwc10 jnthn: so I have hacked MVM_6model_find_method a bit 14:48
it seems that it calls find_method, for "hash", finds a NQPRoutine
and yet still it goes wrong soon after. 14:49
so, random cargo-cult first thought - is something not rooted?
jnthn Is it a concrete NQPRoutine, not a type object? 14:51
nwc10 er, yes, hangon, I don't know that yet.
flags = 16 14:52
I need to look that up
MVM_CF_SECOND_GEN = 16, 14:53
no, it's not a type object
jnthn OK, then at that point things seem legit
nwc10 :-)
"it all goes horribly wrong" is still in the future 14:54
so, I suspect what I now do is
1) take a break, because I'm still a bit fried
2) build master in another tree, with the same extra code and same breakpoint
and see where they diverge
although if I'm feeling lazy I could just set a hardware watchopoint on flags and see where it goes next 14:55
odd. nowhere. it goos "boom" 14:56
ie the backtrace
14:59 zakharyas joined
nwc10 jnthn: it says that it's 88 bytes, and the STable says that it's "NQPRoutine". This means that I can't cast that 88 byets to some C structure and look it it from gdb? 15:02
15:09 SmokeMachine left, SmokeMachine joined
nine nwc10: I'd expect an NQPRoutine to be a P6opaque object. In a debug build you can call MVM_dump_p6_opaque(tc, obj) from gdb to get at its data 15:10
If you're interested in low level stuff, just cast it to MVMP6opaque* 15:11
Also useful: call MVM_dump_backtrace(tc) and MVM_dump_bytecode(tc) to find out where you actually are exactly and get an idea how a value ended up in a register
nwc10 No symbol "MVM_dump_p6_opaque" in current context.
15:13 sena_kun joined
nwc10 and git grep ca't find it 15:13
15:14 Altai-man_ left
nine MVM_dump_p6opaque 15:16
nwc10`(from NQPRoutine) $!do, $!signature, $!dispatchees, $!dispatch_cache, $!dispatch_order, $!clone_callback, $!onlystar=0)
and then I have to know how to do pointer arithmetic to figure out what to cast ot (MVMObject*) ? 15:17
nine Looks like it hardly doesn't contain data 15:18
nwc10 OK. 15:23
I think I'm about to be AFK
nine The dumper works recursively already
The whole point of that is to not need to do the pointer arithmetic :) 15:24
16:09 patrickb left 17:12 Altai-man_ joined 17:14 sena_kun left 18:04 leont left 18:10 zakharyas left 18:53 zakharyas joined 19:13 sena_kun joined 19:14 Altai-man_ left
nwc10 the problem seems actually to be earlier 19:43
in that, master never hits "find_method" for "hash" 19:44
I now have a very chatty set of output from master and my branch
and I'm starting to dig into the diffs
20:29 zakharyas left 20:43 squashable6 left, squashable6 joined
timotimo i think i want to do another attempt at named mallocs/callocs and named fsa allocs, and why not also named spesh allocs just for fun 21:01
moritz my malloc is called Fred! :D 21:03
MasterDuke what do you mean named? 21:04
timotimo mostly turning "thing = (Type *)MVM_malloc(count * sizeof(Type))" into "MVM_MALLOCOBJ(count, Type)", but also there's then a systemtap probe point in there that lets you collect stats and such for allocations "by type" 21:07
er, of course it would be "thing = MVM_MALLOCOBJ(count, Type)" 21:08
there'll have to be yet another piece of probing in the build system for moarvm ... 21:09
at some point we will have to run those in parallel, LOL
21:12 Altai-man_ joined
timotimo or a flag that turns it on 21:12
21:14 sena_kun left 21:15 MasterDuke left
lizmat hmmm down? 21:21
timotimo i think the greater internet is having a bit of a hiccup at this very moment?
lizmat nope, it dropped out of DNS ?
yeah, same for 21:22
moritz I get the IPs and for 21:23
lizmat dig @ A
moritz and the response is SLOOOW 21:24
makes me wonder if the .org TLD servers fail
lizmat dig @ A # works 21:25
moritz a dig @ SOA takes 5 seconds here 21:26
lizmat pinging rba
moritz wtf?
timotimo don't we have "our own" dns server for our domains? 21:27
moritz according to whois, the nameservers are VAL.NS.CLOUDFLARE.COM and CLINT.NS.CLOUDFLARE.COM
and dig @VAL.NS.CLOUDFLARE.COM errors out with "connection timed out; no servers could be reached" 21:28
timotimo: you mean a hidden primary?
timotimo i did get an error from somewhere in cloudflare with a "broken gateway" when loading a pastebin
moritz so cloudflare b0rked?
lizmat VAL.NS.CLOUDFLARE.COM doesn't resolve for me
so something is really borked 21:29
timotimo hearing chatter of cloudflare being down from others
timotimo how do you fall back when cloudflare goes down LOL 21:33
21:34 patrickb joined
moritz 21:42
"Cloudflare is investigating issues with Cloudflare Resolver and our edge network in certain locations." 21:43
seems it's kinda working again 21:44
lizmat yeah, looks like 21:48
22:52 patrickb left 23:13 sena_kun joined 23:14 Altai-man_ left 23:18 sena_kun left