01:49 ilbot3 joined
timotimo not sure if sleep went well 06:08
still getting an explosion when profiling this game i have 06:13
amazing. Mu's "item" method is being specialized into enterspesh, getarg, prof_allocated, set, set, takedispatcher, prof_exit, return 06:15
ah, it used to have a param_sn that got kicked out 06:16
06:31 brrt joined
brrt ohai #moarvm 06:32
so interesting fact; when I do my infinite loop, i have a 'final entry' 06:33
timotimo that feel when your rr session explodes with a malloc problem when trying to MVM_dump_bytecode(tc) ;( 06:39
06:47 domidumont joined
brrt haha 06:50
that feeling when a debugger crashes in general
okay, i'm seeing i officially don't understand 06:51
why compare an object with the flags of the other 06:52
06:54 domidumont joined
timotimo that seems wrong? 06:54
though that ought to mean you know what the fix ought to be? 06:58
nwc10 good *, #moarvm 07:07
timotimo yo nwc10
07:12 brrt joined
brrt kind of wrong, yes 07:13
timotimo break adjusted to free malloc space - why does this happen? :\
the ... f? 07:14
setting "unwindonsignal" off (so that i could debug what's going on) makes it suddenly work instead of crashing with a segfault
oh, now it dies again when i started it from a different place in the code 07:16
just fantastic
i can't reverse-step out of the function i called, so even though i was inside "rr replay" i have to restart the rr replay session :\ 07:18
something's not right with the bytecode dumper helpers i wrote; the place it shows my IP to be at doesn't match up with what the interpreter's doing 07:25
brrt hmmm 07:26
timotimo 9 ops further up there's a findmeth, or 10 ops down there's one 07:27
and the interpreter's currently doing a findmeth op
actually, what if it's looking at the wrong frame on the stack? 07:28
i do see the IP right after an invoke_o
well, it's most likely this frame's bytecode i'm looking at 07:35
i.e. the first few instructions match with the contents of the dump
aha i found where the instruction pointer's at 07:36
i wonder if it's sensible that decont on a low-level null just gives the object itself back rather than asploding 07:42
08:53 robertle joined 09:43 Voldenet joined
jnthn Since there haven't been any complaints yet about gist.github.com/jnthn/8ee6dc06edc7...7ffabb86b5 I figure I'll dig into it somehow :) 10:05
brrt i'm scared but very curious about its reult 10:08
jnthn brrt: What parts of it scare you? 10:11
jnthn is all for trying to reduce risk :) 10:12
brrt well, the multithreaded bit 10:14
but i guess it should be mostly alright :-)
jnthn Ah 10:15
I actually thing it'll be less scary than today
brrt i suppose it is
jnthn In that today *every* thread running might be racing to produce specializations
brrt true
jnthn So every spesh data structure in MVMStaticFrame has to be desinged to cope with multiple writers
Whereas now it'll be single writer, multiple readers
brrt but i mostly use single threaded applicaitons so i never see that 10:16
jnthn True
I think the scheme I suggested for getting predictability will give it, though? 10:17
I'm wondering a bit about taking on gist.github.com/jnthn/8ee6dc06edc7...pe-logging first
Since it's kinda orthogonal-ish to the rest 10:18
Hm, or maybe that's not so wise :)
Hah, no, it ain't
It'll increase the number of deopt points and hose up the recent argument handling changes, I fear 10:19
jnthn makes a branch, digs into the new bits 10:20
Geth MoarVM/spesh-worker: 9c7e6edc7f | (Jonathan Worthington)++ | 6 files
Add MVMSpeshLog representation.

Holds a log to be passed from an executing thread to the spesh worker thread.		 11:01
jnthn lunch & 11:04
brrt nods, that scheme ought to work 11:11
11:18 brrt1 joined 11:38 nebuchadnezzar joined
JimmyZ jnthn: replied your gist.github.com/jnthn/95d0705d225e...aa337624b, though maybe I am wrong 11:44
jnthn That's somewhat orthogonal to box/unbox 11:48
Though very much related to escape analysis
timotimo wow, 32bit for used and alloc :) 11:50
jnthn Not sure how long the logs will want to be :) 11:51
timotimo it's 2x 64bit for every entry
m: ((2 ** 32 - 1) * 8) / 1024 / 1024
camelia WARNINGS for <tmp>:
Useless use of "/" in expression "/ 1024 / 1024" in sink context (line 1)
jnthn Really? :)
timotimo m: say ((2 ** 32 - 1) * 8) / 1024 / 1024
camelia 32767.9999924
jnthn used/limit are in terms of entries, not in terms of bytes :) 11:52
timotimo so, 32 giga-entries?
yeah
32 gigabytes in entries per log, i mean
m: say ((2 ** 16 - 1) * 8) / 1024 / 1024
camelia 0.4999924
timotimo or half a gigabyte of log data per thread if we went down to 16
JimmyZ wonders what jnthn's thought about Allocation Sinking Optimization compare to other EA Algorithms :P
jnthn We could, though given there's one log held per thread :P 11:53
timotimo though of course there's alignment that might force used + limit to be 64bit wide anyway
jnthn JimmyZ: It's not clear that the AS thing isn't tied to tracing
timotimo: Yeah, it seems like an odd place to try and save a few bytes :)
JimmyZ jnthn: oh, yea
timotimo true that
JimmyZ yeah
jnthn On a 64-bit system an entry is actually 24 bytes 11:54
kind is 4 (alignment makes it being less not worth it)
id is 4
Then the biggest union branch is 16 11:55
(entry, 2 pointers)
timotimo wow, how'd i miscount that?
jnthn Not enough lunch? Too much lunch? :)
timotimo i've only had breakfast and a nap 11:56
i think now the cat is demanding attention 11:59
12:18 AlexDaniel joined
timotimo okay that was almost an hour of neck-scratches and bellyrubs 12:46
Geth MoarVM/spesh-worker: 465451fab3 | (Jonathan Worthington)++ | src/core/instance.h
Try to organize instance.h

It has grown a huge amount of state over time. This attempts to group the related pieces in some way, to get a better sense of what's in there.		 12:47
timotimo haha there was still a reference to APR in there 12:48
jnthn Yeah :P 12:49
Finally I can see what state spesh actually has :)
heh, tempted to do a simialr thing for threadcontext.h too 13:08
Having just started to do it - yes, 'cus I think I found a dead field and code to go with it too :) 13:13
timotimo the fun i'm having ATM is things asploding in a non-speshed frame, though of course the bad data could come from a wrong-speshed frame 13:17
Geth MoarVM/spesh-worker: 6ddfcffb12 | (Jonathan Worthington)++ | src/core/threadcontext.h
Organize the mass of state in threadcontext.h some

So it's easier to see what's in there and what key subsystems are using.		 13:20
MoarVM/spesh-worker: ac8450a6b5 | (Jonathan Worthington)++ | 3 files
Add/mark spesh log queue and per-thread spesh log

Currently these are not yet initialized.		 13:30
14:05 SourceBaby joined 14:26 colomon joined
Geth MoarVM/spesh-worker: 0f1de2d36d | (Jonathan Worthington)++ | 4 files
Remove now-unused cur_usecapture field.
 14:35
MoarVM/spesh-worker: 07b722958b | (Jonathan Worthington)++ | 6 files
Start specialization worker thread at startup.
timotimo sigh, the internet connection here keeps dying 14:49
interesting, we don't keep a handle on the spesh thread in the instance object after all? 14:50
i suppose we don't need any other handle than the queue 14:53
jnthn timotimo: I was trying to think of what I'd need it for and couldn't :)
If we get a reason to ahve it we can add it :) 14:54
timotimo yeah, i can't come up with anything either
Geth MoarVM/spesh-worker: 3fb2d40d2c | (Jonathan Worthington)++ | 7 files
Set up initial per-thread spesh log, if enabled.
 15:15
MoarVM/spesh-worker: 937dcb2cb0 | (Jonathan Worthington)++ | src/spesh/worker.c
It helps to actually run the created thread.
timotimo heh. 15:19
man you runed it!
15:21 brrt joined
brrt he wants to JIT compile a specialization selector 15:26
okay, we can do that
but
if we're going to go that way 15:27
why not compile all specializations in a single frame
oh, i know why
because we might not be able to compile all specialized versions
timotimo whenever we add another specialization, we'd have to recompile the whole thing, no? 15:35
that could perhaps be expensive
and re-use of registers and such will have to be A Thing, as opposed to how we do inlining where we just put the whole work of the inlinee at the end of the inliner's work area
jnthn It has occurred to me that we might be able to shortcut quite a lot of the invocation work, though :) 15:37
brrt i'm listening
jnthn And have the selector end up just jumping into the JITted target code
Frame creation has quite a few conditionals and calculations and so on
timotimo ah, yes, interesting 15:38
making invocation cheaper is something we'll all benefit from
jnthn Bump-the-pointer'ing the CallStackRegion for example should be easy to compile
And sometimes that's all we'll have to do
15:56 brrt joined
brrt nods 15:57
and suggest that the expr jit should make this kind of thing easy
because it introduces a layer that the 'C' code can work with to make machine code 15:58
could also be done in dynasm, for sure
…. oh, wow 16:25
how even does this happen
my label is off by one....
16:47 AlexDaniel joined 17:03 robertle joined
timotimo if i need to make a branch for advanced moarvm-under-gdb stepping functions (step in and out of moar-bytecode-level funcitons and such), i'll probably call it here_come_the_hotsteppa 17:35
jnthn :P 17:39
18:16 domidumont joined 18:29 zakharyas joined 19:09 buggable joined
samcv \o/ yay it's working 20:38
and seems to be working even when i dump all the data into it. not just a partial run
and also work when there's collation elements attached to a non-terminal node
HelpBot[perl6] \o/ 21:14
21:16 lizmat joined
samcv and luckily the code has less paths now as well. so that's a ++ as well :) whenever doing recursion the more code paths really can screw you over 21:18
now it just has ternary operators and maybe one if condition period 21:20
for the code that creates the c code
21:58 nebuchadnezzar joined

Powered by Cro and the Raku® Programming Language.

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

Thank you!