02:13 vendethiel joined 04:10 tokuhiro_ joined 07:38 FROGGS joined
nwc10 good *, #moarvm 07:55
09:11 zakharyas joined 10:48 brrt joined
brrt good * 10:49
i've a question 10:50
for symmetry reasons it would be.. reasonable to add a 'zero label' at the beginning of the jit code
well, let me first explain the challenge
lizmat brrt nwc10 o/
brrt \o lizmat 10:51
i've decided on the following labeling strategy; all basic blocks receive a label that is equal to their basic block index; the graph end recieves a label equal to the number-of-basic-blocks (hence 1+ largest basic block); 10:52
only before-instruction labels exists, and they are managed in an associative array, whereby the index + the instruction label offset equals the label number 10:54
internal expression labels are calculated after that still
anyway, because the first basic block is always a NOP, it is simply skipped during JIT graph building 11:01
(first bb of a spesh graph, that is)
but this bb has index 0; hence no basic block has index 0; hence this label is not initialized, hence a crash 11:02
(label 0 is used as the initial entry label for a newly created frame)
timotimo brrt: you could just put nops in the code? 11:03
brrt i though it would be a sensible idea to have label 0 as the 'start of graph' label, except that it is of course completely redundant
anyway, lecture &
timotimo good lecture, brrt
lizmat .tell jnthn I think we have a problem with nqp::readlinechompfh : if the last line in a file is empty, nqp::eoffh is 1 and nqp::chars=0 11:10
hmmm...no yoleaux here ?
.tell I wouldn't expect eoffh to be true until the next read 11:11
timotimo yeah, no yoleaux in here
11:23 camelia joined
lizmat $ 6 'spurt("one","\n"); say open("one").get' 11:27
(Str)
11:27 camelia joined
nwc10 brrt: yes, wingolog.org/archives/2015/11/03/tw...mentations is interesting. And seems to have created a sideline in (C)Python speed optimisation snark. For reasons I don't know why 11:39
FAT Python looks interesting enough - as in, I'm curious to see how the experiment pans out: article.gmane.org/gmane.comp.python.devel/155300
given that "no plan survives contact with the enemy"
eg blog.kevmod.com/2014/04/pyston-faq/ -- PyPy uses a tracing loop JIT, and Pyston uses a method-at-a-time JIT. A tracing JIT like PyPy can be impossible to beat when they hit their fast path, but my hope is that a method-JIT like Pyston could hit its fast path more often and more reliably, especially for larger codebases, which will end up being a win even if the fast path is not quite as fast. 11:41
jnthn timotimo: We actually compute the NFC/NFG "together" rather than producing all the codepoints as an intermediate step, and encodings are free to not send stuff through the normalizer; that's what the utf8-c8 decoder will do. 11:52
timotimo oh! 11:53
that's good!
... AFK for a bit 11:54
oh, also: github.com/MoarVM/MoarVM/issues/149 this issue is about a "die" in a signal handler not causing the program to exit. i think it shouldn't, but i wanted input from someone else before i reject the addendum to this ticket.
jnthn timotimo: That's not a MoarVM issue any more 11:57
And there's already an RT about die in a tap getting swallowed 11:58
dalek arVM/utf8-c8: 8c0623f | jnthn++ | src/strings/utf8_c8.c:
Factor out repetitive resizing code.
 12:20
arVM/utf8-c8: e97782b | jnthn++ | src/strings/utf8_c8.c:
Flush normalizer upon an undecodable.
 12:28
brrt nwc10, oh, backloggings 12:41
dalek arVM/utf8-c8: 065cd60 | jnthn++ | src/strings/utf8_c8.c:
Fix reject loop to not lose bytes.
brrt nwc10: i can kind of explain the wingolog snark, i think 12:42
nwc10 oh? The comment snark seemed to be from people other than the blog owner 12:43
brrt the cpython core development team has decided not to add features that improve performance by x% if they increase complexity by >x%
oh, really?
i didn't read the comments
nwc10 there was somewhat snark in the blog
and then similarly toned comments from at least one person 12:44
brrt well, it is a reasonable question; it's almost as if the cpython team doesn't care (gasp)
and the cpython team *doesn't* care, and this is within their rights to do so
nwc10 and *yes*, it did all seem to agree - that the CPython folks aren't fretting about performance
brrt i liked the expression 'c extension furnace' 12:45
nwc10 me too 12:46
loved that one
arnsholt Yeah, that was a brilliant turn of phrase
jnthn Choosing to compete by being the solid/stable/reliable choice is reasonable. 12:47
Many people *will* value that over performance
Part of the challenge of MoarVM development is that we need to both provide reliability and push the Perl 6 performance boundary too. 12:48
nwc10 Choosing to avoid painting yourself into a maintenance corner in the middle future also makes sense
in the same vein, it's important to have "rigidly defined areas of doubt and uncertainty" otherwise you have no wiggle room to fix any bugs ever 12:49
"but the existing implementation did that, and I relied on it"
arnsholt Yeah, Python suffers a bit from that, I think 12:50
jnthn We've yet to see how well "if it's not covered by the test suite it's our wiggle room" works out for Perl 6. :) 12:52
arnsholt Yeah, obviously =) 12:53
But at least there's some consideration given to the potential existence of several implementations
We've got that going for us, which is nice =)
nwc10 brrt: it's really hard to assess 'increase complexity by >x%' 12:56
brrt yes
nwc10 I agree that it's a legitimate choice. It's just that it's very hard to try to be objective about it
brrt well, it is a costly choice, too 12:57
on the other hand....
the other side of it, is that the JIT has seen a somewhat oscillatory process of complexity 12:58
i.e. I add something to make something work, then I change it to be more understandable, then I add new features, then I rework them
that is basically the only way I can keep going
if you insert a barrier at >x% complexity, then you may never cross it 12:59
because it tells people 'don't bother doing that, it is too complex'
so, there is an important signal value in not wanting complexity, aside from the effect of reducing complexity 13:00
and as an aside, the guile vm looks actually increasingly promising, i think 13:01
timotimo: do you recall who it was that asked for illumos? 13:04
timotimo it was konobi 13:07
brrt hmmm... doesn't seem to be online now 13:10
i would kind of need a good testcase for this 13:11
it'd be really stupid if illumos support would fail because of this minor issue in minilua 13:12
on the other hand
we can configure to disable the JIT for illumos
timotimo why is konobi not online? at least in the nicklist 13:13
brrt oh, he is 13:14
i was looking wrong
timotimo that's why i relied on my tab key
brrt is using webchat because of the great it department of my $study-office 13:19
timotimo %) 13:20
brrt i can't listen to internet radio, either :-( 13:22
(first world problems much... :-P)
13:23 stmuk_ joined
timotimo that's a productivity reducing problem, IMO 13:25
well, it can be. depending on the person in question
13:26 kjs_ joined
brrt happiness reducing problem, i think 13:29
also, reducing stimuli problem in a shared office 13:30
dalek arVM/utf8-c8: 1650c8b | jnthn++ | src/strings/utf8_c8.c:
Fix off-by-one on the undecodable.
 13:40
arVM/utf8-c8: 58f2bbb | jnthn++ | src/strings/utf8_c8.c:
Implement UTF-8 C-8 encoding.
brrt (also the phrase 'powerful, bitter C++ wizard' - rather strong image) 13:58
14:35 brrt joined 15:08 colomon_ joined
dalek arVM/utf8-c8: 96fafa0 | jnthn++ | src/strings/utf8_c8.c:
Implement UTF-8 C-8 streaming decode.
 15:14
jnthn Phew
nwc10 in that, you think that it's done?
jnthn The encoding, yes 15:15
Now I "just" need to update various places to use it
timotimo much triage
jnthn I also can't really write tests for its effectiveness on Windows
Because Windows was sensible enough specify the encoding for file system entries 15:16
brrt wait, what
jnthn I mean, the encoding I can test
brrt filesystem entries (file- and directorynames) are encoded in windows? :-o
hmmm
brrt wonders.. and wonders
and sees potential for abuse 15:17
jnthn Well, hmm 15:19
"Use any character in the current code page for a name, including Unicode characters and characters in the extended character set"
timotimo so ... i wonder why loop { start { say "hello" } } hangs. should i spend time investigating that? 15:20
jnthn timotimo: I've no idea why it would
timotimo (that's issue 140 on github)
lizmat perhaps it doesn't get a lock to actually cue ?
I mean, start doesn't actually cue, does it ? 15:21
jnthn It cues
Well, it calls Promise.start({...}), which does
brrt cue? 15:22
lizmat so it pushes it to $!ququq
$!queue
timotimo it takes its sweet time to actually hang ... 15:23
lizmat does anything reading from the queue get a chance in that loop { start {} } setup ?
timotimo or perhaps i'm failing to reproduce that right now
sure, the code blocks get run all the time
now it hangs!
jnthn lizmat: Well, it's another thread that's reading
timotimo: Neat. Now get backtraces of the threads (C level should do) under the debugger :) 15:24
timotimo oh. why did i ctrl-c? :| 15:25
(because another snippet in that issue uses sleep 0.01 instead of say and is said to segfault)
it doesn't segfault here, but surprisingly its memory usage grows
nwc10 jnthn: it seems to break 15:26
timotimo stopped it just above a gig
nwc10 not ok 96 - nqp::escape
dalek arVM/utf8-c8: 3f77ddc | jnthn++ | src/strings/utf8_c8. (2 files):
Add "encode to C string" function for UTF-8 C-8.
jnthn nwc10: "it"?
nwc10 origin/utf8-c8
yes, sorry, return key faster than me
jnthn nwc10: Suspect missing patches from master 15:27
nwc10 is that one of those "excuses" like: 15:28
15:28 <dipsy> Your excuse is: clock speed
:-)
lizmat so if I understand C-8 correctly, it creates 256 synthetics for each possible byte value ?
nwc10 well, 128, if I understand it, because it's only the octets 128-255 15:29
timotimo gist.github.com/timo/e718153e22177958174b - fun! 15:30
lizmat wouldn't that be usable for allowing regexen on binaries, given the right syntactic sugar ?
brrt timotimo: you might want to run that with MVM_JIT_DISABLE=1 15:31
and see if that changes anything, anyway
timotimo one thing about regexes on binaries is that we'd like to switch interpretations between 8 and 16 and 32 bits and such
brrt: can do!
brrt :-)
timotimo i forgot about the backtrace screwyness we get ... 15:32
lizmat I think a lot of people would already be very happy with 8 bit :-)
timotimo it looks a lot like a case we didn't yet cover in the GC orchestration code
brrt: any reason to go "bt full" on this? 15:33
brrt other than why not, i'm not sure
timotimo there's a second file in there now
brrt i'm surprised by how many threads there are
oh, it's a loop, duh 15:34
timotimo brrt: because of LABELS = {0x7ffff7afd0fe <MVM_interp_run+142>, 0x7ffff7b0dd31 <MVM_interp_run+68801>,
brrt oh, well, then not
anyway, my spider sense is thinking OSR
probably wrong, but worth a shot
timotimo now there's a third file in there 15:37
helpfully that file is right on top
jnthn nwc10: No, there actually is a patch in master about escape that landed after I started the branch :) 15:38
timotimo could just be that shift of ConcBlockingQueue has to mark the thread as "blocked"
jnthn timotimo: Ah, it doesn't? 15:39
It should
nwc10 people other than you commit stuff‽
jnthn I know, shocking!
timotimo hm, actually
it seems like it does
jnthn The project's at risk of having a community or something :P
timotimo because inside the thread that shows that particular variable, num_threads is 2
thread 15 is doing some atomic integer faffing about inside orchestrate.c 15:40
thread 1 is in gc_enter_from_allocator and has num_threads = 2
brrt we shouldn't let the bus factor get too high :-o 15:41
timotimo brrt: makes it more expensive for the government to shut our project down
thread 1 spins on if (threads && threads != last_starter) ... while (MVM_load(&tc->instance->gc_start) > 1); 15:42
thread 15 is inside the while (MVM_load(&tc->instance->gc_start)) 15:43
oh, would you look at that 15:45
print tc->instance->threads
0x0
who does that?
jnthn o.O 15:46
timotimo for some reason there's an if(threads && ...) in there 15:47
so ... do we expect tthat to be null in some cases?
dalek arVM/utf8-c8: 3f40008 | jnthn++ | src/ (11 files):
Use UTF-8 C-8 encoding in OS/IO interactions.
timotimo like, if there's only the one main thread?
timotimo is definitely not on the GC orchestration bus 15:48
jnthn $ env - ACME=$'L\xe9on' ./perl6-m -e 'say "hi"' 15:49
hi
\o/
timotimo neato! 15:50
m: say chr(\xe9)
camelia rakudo-moar ea87a8: OUTPUT«===SORRY!=== Error while compiling /tmp/2b1CTawTcg␤Undeclared routine:␤ xe9 used at line 1␤␤»
timotimo m: say chr(0xe9)
camelia rakudo-moar ea87a8: OUTPUT«é␤»
timotimo m: say uniname(0xe9) 15:51
camelia rakudo-moar ea87a8: OUTPUT«LATIN SMALL LETTER E WITH ACUTE␤»
timotimo ah. damn this font.
brrt timotimo: we're distributed over multiple government areas already, so that makes it pretty hard for one government to shut moarvm down 15:54
except ofcourse by the TPP agreement
16:02 kjs_ joined
dalek arVM/utf8-c8: 43377bb | jnthn++ | src/strings/utf8_c8.c:
Be a little smarter about undecodables.
 16:03
timotimo there is already a single world government. it's the one run by the lizard people 16:10
brrt lousy government at that, with so many civil wars, then 16:12
timotimo yeah 16:13
well ...
really depends on what metric you apply
they clearly have something different in mind than "world peace and the continuation of the human race"
that's not "typical lizard people stuff" 16:14
bbiab
jnthn Darn, that was a silly bug 16:18
dalek arVM/utf8-c8: e5a39f6 | jnthn++ | src/strings/utf8_c8.c:
Fix forgotten function call rename.
 16:19
jnthn $ echo 'say(42)' > $'L\xe9on' && ./perl6-m $'L\xe9on' 42 16:20
fail paste, that 42 at the end shoulda been on another line
But anyways, it works.
brrt \o/ 16:21
dalek arVM: 1beea0b | jnthn++ | / (9 files):
Work in progress UTF-8 Clean-8bit encoding.

For allowing us to round-trip filenames and so forth. Only partially implemented so far. ee75b65 | jnthn++ | src/strings/utf8_c8.c: Tweak comment to be lik-er.		 16:39
jnthn Dear dalek, that was a merge :P
16:40 dalek joined
17:29 kjs_ joined
dalek arVM/even-moar-jit: 7d88c3e | brrt++ | / (10 files):
Refactor labelling

Now all basic blocks receive a label equal to their basic block index. Because we skip the first basic block, the 'zeroth' label is available for the graph start, and the end-of-graph label is simply equal to the number of basic blocks (hence 1 greater than the latest basic block). Individual instructions are labeled, if necessary, using an associative array as before. Internal expression labels are computed during compilation based on the offset.		 17:35
17:36 brrt joined
brrt of course, the real work is now porting the labeling stuff to the epxression jit 17:38
i'm wondering if i can try to make the interface between these things shared 17:39
jnthn brrt++ # the scheme sounds sane 17:41
jnthn bbl
brrt :-) 17:42
17:44 brrt joined
JimmyZ jnthn: I saw there was a gist about type coercion, I lost the url ... 17:45
17:47 kjs_ joined
TimToady a datapoint on the non-complexifiying of Python; at Hackers Guido was showing off $9 Python board that they're gonna distribute a million of to British schoolchildren 18:01
I've always emphasized community, but Guido has usually aimed more for institutionalizing, which I figured out when he went to work for NIST :) 18:03
nwc10 what's "Hackers" in this context? I have a suspicion that it's pretty hard to Google
japhb has to mentally recite "If you don't have anything nice to say, don't say anything at all" WRT institutionalized Python 18:04
nwc10 the board will be this, unless I'm confused: en.wikipedia.org/wiki/Micro_Bit
18:04 colomon joined
nwc10 my employer had been at risk of becoming an institution. And those that could survive the environment, institutionalised. 18:05
Fortunately, regime change happened.
japhb My wife once had a contracting job that was so stressful that people literally left the team via heart attack. 18:07
(She was sufficiently mobile to leave of her own volition first, but sheesh.) 18:08
nwc10: If the specs on the wikipedia page are correct, the I/O controller on that board is significantly more powerful than the CPU. 18:09
TimToady nwc10: the 31st Hackers Conference, which was this weekend 18:11
18:11 kjs_ joined
[Coke] still can't find a url even with that. (one major hit, but it's from 2014) 18:12
nwc10 me neither. :-)
that probably shows how cool it is 18:13
TimToady it's invite only
and very off-the-record 18:14
no photos or recordings, by policy 18:15
[Coke] You're probably violating the first rule of hackers conference right now. 18:20
arnsholt Sounds vaguely like Chatham House rules (or whatever the name for it was) 18:21
TimToady [Coke]: undoubtedly :) 18:25
I don't think the project itself is secret, but perhaps the fact that Guido mentioned it at Hackers is :) 18:26
18:27 FROGGS joined
nwc10 no, not secret now. eg pyfound.blogspot.co.at/2015/07/bbc-...nched.html 18:28
www.bbc.co.uk/mediacentre/mediapacks/microbit
and most intestingly ntoll.org/article/story-micropython-on-microbit 18:29
TimToady "batteries not included" :)
timotimo moarvm has no chance to run on a slim device :| 18:30
well, moarvm itself does
rakudo ... not so much 18:31
japhb wonders about nqp 18:52
18:52 Peter_R joined
japhb Because for the embedded case, that would feel about right. 18:53
18:54 vendethiel joined
TimToady JimmyZ: you were looking for gist.github.com/anonymous/8efdab238ead35b1d4fa maybe? 19:02
19:13 sivoais joined 19:25 sivoais joined 19:38 TimToady joined
jnthn TimToady: I don't recall you ruling if you agreed with that gist or not ;) 20:00
timotimo jnthn: i was thinking about frames that may have multiple osrpoints and such; wouldn't it make sense to only kick out the osrpoint(s) that actually get hit? because of our logging and stuff? 20:07
and ... would that work at all? causing an already spesh'd frame to be spesh'd again? 20:08
and if it's not worthwhile to keep osrpoints around, perhaps we want to put in something that'd re-spesh after 10x as many rounds, so that we catch many more spesh'd callees on the second try 20:09
jnthn How common is having multiple osrpoints? 20:11
timotimo good Q. i was hoping to hear your intuition on it, but i could actually measure it 20:12
jnthn My suspicions is "not highly common"
timotimo i was thinking something like a little loop to prepare data, then a big loop to do stuff
and then the first loop would "steal" the logging phase 20:13
timotimo does too much thinking with too little measuring
it's just that i have no answer for the "we spesh outers before our callees get speshed" situation and it bugs me
20:32 kjs_ joined 21:07 Ven_ joined 21:08 Ven__ joined 21:12 kjs_ joined 21:15 synbot6 joined 21:21 Ven_ joined
hoelzro jcdav.is/2015/11/09/More-JVM-Signal-Tricks/ 21:35
neat trick!
lizmat m: "\r\n".ords.say 21:43
camelia rakudo-moar aaf9fc: OUTPUT«(13 10)␤»
21:51 Ven_ joined
konobi okay... i checked with the libuv folks... it seems that the compile error is specific to this build 21:57
ugh... looks like an illumos glitch 22:05

Powered by Cro and the Raku® Programming Language.

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

Thank you!