00:01 bisectable6_ joined
samcv timotimo, that would be nice :) 00:05
01:29 colomon joined 01:40 colomon joined 01:41 colomon joined 01:52 ilbot3 joined 03:05 pharv_ joined 03:44 stmuk joined 03:46 Util joined 06:15 lizmat joined 06:22 lizmat_ joined 06:38 lizmat joined 08:25 zakharyas joined
jnthn moarning o/ 09:10
09:12 lizmat joined
dogbert17_ morning, why do I get the impression the everyone is in Amsterdam 09:17
09:17 zakharyas joined
jnthn I dunno, maybe most of 'em are 09:18
dogbert17_ so what's on the agenda today, closures and/or tuning? 09:19
jnthn Think I'll make an attempt at the closures thing 09:21
Though running a little short on time
dogbert17_ is it a straightforward fix or one requiring a lot of thought and coffee? 09:23
jnthn It's not really a fix
It's more trying to remove a limitation of inlining
It's tricky
dogbert17_ and so can take an unspecified amount of time 1 .. * 09:24
jnthn Well, I've got a rough plan that I think might work
First step is to try and teach deopt to find the correct code ref via a register
So uninlining would be able to reconstruct things correctly 09:25
Which today doesn't matter
The basic idea is that we add a spesh op that can resolve a lexical relative to the outer of a code-ref held in a register 09:26
otoh since an MVMFrame is a collectable, we could extract the outer frame, but maybe still need the coderef for deopt anyway 09:27
dogbert17_ was temporarily called away, a disadvantage of $work 09:51
any wild theories as to performance increases with this new opt? 09:52
09:54 brrt joined
Geth MoarVM: 9f12f0c1a6 | (Jonathan Worthington)++ | 7 files
Start storing/using static frame in inline table.

In the vast majority of cases, the code ref was only used as a way to get at the static frame, so now just store and access it directly. This prepares the way for starting to store the code ref instead by a reference to the register holding it.		 09:55
brrt good hi *
jnthn dogbert17_: Well, inlining can be a notable win
timotimo yo brrt 09:56
jnthn o/ brrt
timotimo when's the jit talk?
brrt tomorrow
jnthn How's Amsterdam?
timotimo OK!
brrt i'm not there yet
:-)
jnthn Oh!
brrt long story
but i'll be there in 2h :-)
yesterday, it was kind of rainy
.tell nwc10 you should be able to build nqp by now 09:58
yoleaux brrt: I'll pass your message to nwc10.
10:12 lizmat_ joined 10:16 lizmat joined 10:20 brrt joined
timotimo so i was wondering how to implement getting the remote peer from IO::Socket::Async's listen functionality 10:21
jnthn timotimo: Just pass it as an extra argument or two to the callback that fires per incoming connection 10:22
timotimo that sounds good, that'll get us the feature without a new op 10:23
jnthn Right :)
Geth MoarVM: 30ff9d5147 | (Jonathan Worthington)++ | 5 files
Use register holding code ref for deopt purposes.

This means that we will deoptimize to the correct closure clone. This also prepares the way for an attempt at inlining of closures.		 10:28
jnthn Well, seems my preparatory work went without hitches.
nwc10 it's a trap! :-) 10:29
yoleaux 09:58Z <brrt> nwc10: you should be able to build nqp by now
nwc10 (I'd found that already last night)
10:31 lizmat joined
brrt is now taking the express train to amsterdam 10:34
timotimo can't pay attention to talk and code att he satem tim 10:46
hmmm ... nqp-debug-m ... 10:49
10:49 brrt joined 10:55 brrt joined 11:09 markmont joined
jnthn Well, got a patch that works for the NQP build/test under spesh stressing. Explodes during CORE.setting compilation alas 11:27
Lunch, will investigate after
11:38 markmont joined 11:48 lizmat joined
jnthn back 12:06
Of course, the CORE.setting build is the worst place to hunt a problem down, so I'll see if it'll built without nodelay/blocking and then see if any tests fail with it 12:07
nwc10 or watch Damian live... 12:08
jnthn I get to do that in a couple of weeks :)
nwc10 he seems to have just blown some folks' minds. 12:09
12:12 colomon joined
jnthn :) 12:15
jnthn found a simpler test to investigate
Ah, a rewrite thinko 12:30
yay, and with that fixed it built Rakudo with spesh stressing, and passes make test too 12:35
Next for spectest with spesh stressing 12:36
12:41 lizmat joined
jnthn Just a few busted 12:47
otoh yesterday there was one busted and it took all afternoon to work out why... 12:48
12:48 colomon joined
dogbert17_ jnthn: these ones? S03-junctions/autothreading.t, S05-modifier/ignoremark.t, S05-modifier/samemark.t, S32-io/IO-Socket-INET.t and integration/eval-and-threads.t 12:50
jnthn No
dogbert17_ :(
jnthn But it turnsout
They are I think the same tests I regressed when I hadn't JITted new ops a couple of days ago 12:51
And it turned out it was an issue going way back
So I think I've just stumbled on that again
Time to write some assembly I guess...
dogbert17_ making brrt unemployed :) 12:52
jnthn I'm not *that* good at assembly :P 12:57
Also the hard part of what brrt++ is doing is figuring out how to generate good assembly, whereas I at the moment just need to write a block of it for the lego JIT 12:58
nwc10 at this time any other JITs that might exist are merely the figments of someone else's imagination/on a branch/wrapped in an SEP field
and IIRC he's on a train so won't notice for a while :-) 12:59
jnthn Ooooh. 13:02
OK, looks like I've some more fix to fix 13:04
*stuff to fix 13:05
But with the new ops JITted and a tweak to the inline threshold, we get another bit of movement on the read a million lines benchmark.
dogbert17_ figuring out what is good assembly must be much harder today than in, let's say, the 6502 days 13:09
jnthn Well, my assembly at least passes spesh stressing on NQP build/test and Rakudo build/test 13:12
dogbert17_ coool
jnthn Now for another spectest run to find out what I need to fix.
13:12 lizmat joined
geekosaur arguably it was harder back then, because you had to think more. 13:20
*especially* on 6502 where, if you wanted really good asm, you had to decide which indexing mode to use (which determined whether/how you used the index registers) 13:22
that said, the two worst cases are also in the past. RISC, and VLIW (the late unlamented Itanium)
dogbert17_ interesting: I was convinced that it had actually become harder, having to think about caches, branch predictors, pipe stalls etc 13:25
geekosaur those are issues, yes, but more meta-issues in some sense. 13:26
and even there, RISC was nastier --- delayed branch slots were supposed to make that easier, but instead made it harder in many cases 13:27
by "meta-issues" I mean that to do it right, you have to think about more than just the current asm, you have to think about the program context it is invoked from 13:28
13:29 brrt joined
dogbert17_ very educational, fortunately for me, the last time I had to do anything with assembler the cpu of choice was the 68000 13:29
geekosaur so they're more matters of *global* code generation/optimization. which is one reason LTO is becoming a thing; it has a more global viewpoint and can do a better job with that kind of optimization
brrt .tell nwc10 i've already long since arrived at the conference 13:30
yoleaux brrt: I'll pass your message to nwc10.
nwc10 "I'm here all week"
yoleaux 13:30Z <brrt> nwc10: i've already long since arrived at the conference
brrt :-) 13:31
nwc10 (but not necessarily awake)
brrt also, making myself redundant is very much the endgame
dogbert17_ LTO, is that what gcc is doing?
geekosaur gcc and clang both support LTO but not by default
brrt i think msvc dooes it by default
it makes build times really really bad
geekosaur ("link-time optimization")
brrt it's also pretty cool optimization 13:32
dogbert17_ do we, i.e Perl 6, use it
geekosaur no, and I don't think it realistically can
brrt dynamic optimization is in a way similar 13:33
in that it works on the rprogram in its entirety
but only on the parts that are actually running
that's usually a pretty good tradeoff of resources
jnthn We can inline across compilation units already, which I guess is one of the things that LTO does with C code 13:34
afaik, prior to that, inlining was only within a file?
geekosaur C99 has extern inline functions 13:35
but it has to be explicitly declared as such (and in a header file so it can be #include-d) 13:36
brrt jnthn, i'm curious why your branch breaks when the JIT can't JIT the right things 13:37
that's worrisome
13:37 lizmat joined
jnthn brrt: I don't know. 13:38
brrt: But I traced it back to some time before my recent work 13:39
So I think it's one of those spesh bugs that's long existed 13:40
brrt hmmm
and JITting it fixes it?
jnthn I've found/fixed many of those
brrt that
jnthn Yeah
Which is *really* odd
brrt which is really scary
jnthn That too
brrt because to me, it sounds like the JIT might cover up some bits 13:41
jnthn First test I tried to bisect with the spesh limit didn't lead anywhere good. Second one has led to something that looks more sensible 13:50
In common they seem to have exceptions
14:00 zakharyas joined
timotimo oof, so handlers and maybe how inlining treats handlers 14:11
Geth MoarVM: 55f850ca41 | (Jonathan Worthington)++ | src/spesh/graph.c
Another place needing deleted handler handling.
 14:12
MoarVM: af24ab825a | (Jonathan Worthington)++ | src/spesh/inline.c
Can never inline frames declaring state vars.
14:15 markmont joined 14:17 brrt joined
jnthn ah good, I got a golf 14:18
Ahhh. 14:27
OK, DEPRECATED assumes $bt.next-interesting-index will always return a value, in one case
But because we're inlining more aggressively these days, that doesn't actually happen
Backtraces could really do with inline fixups 14:28
14:35 lizmat joined 14:40 brrt joined
jnthn oh man...1 line fix. 15:07
From a thinko from 2+ years ago
Or so it seems
15:10 brrt joined
japhb Amazing how long those things can lie dormant ... 15:17
jnthn Indeed. otoh, we inline quite a lot more than we used to 15:19
Or, the other way round, we didn't have so many inlining opportunities before. :)
japhb Heh. 15:20
Let me know when to do a Rakudo rebuild, and I'll try again with my collection of what seem to be sensitive module installs.
Geth MoarVM: c8d902a0ed | (Jonathan Worthington)++ | src/spesh/inline.c
Correct handler order when inlining.

We duplicate the exception handlers in force when inlining. It's true that order matters, and innermost first is searched for. Unfortunately the effort to do the right thing about that was misguided here, since we're copying from a table that's already sorted, so should just leave it in the order it was already in.		 15:29
15:35 zakharyas joined
Geth MoarVM: c663342b48 | (Jonathan Worthington)++ | 9 files
Implement inlining of basic lexical lookups.

This means that we can now inline closures that look up symbols they close over. It works by introducing a new pair of lookup ops that find the code ref in a register and use it to find the outer. After JIT, this turns out to be as many dereferences as in the non-inlined case.		 15:46
timotimo whoops 15:48
jnthn Oh, maybe one more instruction actually...forgot the initial register access.
timotimo but inlining will still give us the power to optimize more stuff
jnthn timotimo: The "as many dereferences as" is a good result
It means that the machinery to make the inlining work doesn't make the lexical lookups more expensive. 15:49
timotimo oh!
yay :)
japhb "only as many"? 15:50
jnthn "only as many" might have been clearer 15:51
Geth MoarVM: e50eb617db | (Jonathan Worthington)++ | src/spesh/inline.h
Tune the inline size a bit.

Based on looking at some benchmarks. The original bytecode size often shrinks notably during the specialization process in Perl 6 code, as the argument prelude gets gutted. We do check the size after, but it will then include any nested inlines, which made us more reluctant to do multi-level inlines. This could likely be tuned further; for now this seems a win.		 15:54
jnthn With this lot, I measure us slightly beating Perl 5 at the "read a million lines with UTF-8 and add up the chars" benchmark 15:55
With the bonus that we're counting graphemes, not codepoints.
Something's a tad off with the profiler though 15:56
The wallclock times and the spesh log both show that the loop body is 100% inlined
Like, the + operator and .chars call are inlined into the loop body, and the loop body is inlined into the unit frame. 15:57
But for some reason that isn't shown
It shows the + and chars, but not the pull-one 15:58
uh...I mean the loop body
Oh but...maybe the extra profiling instructions push it over the inline limit? :)
That probably will be it :P
15:59 zakharyas joined
jnthn Anyway, hopefully this will have other similarly positive effects :) 16:00
Sadder news is that the budget for doing this stuff is about spent, so that'll be the last major new opt for now. There's a little left, while I'll use to see if the named arg alias regression that got RT'd is still a thing and hunt it down. And maybe a little further tuning if time allows. 16:06
Really happy to have got to the point of having this kind of inlining though :) 16:07
Time to go cook dinner. Yes, it is Indian today :) 16:12
timotimo \o/ 16:18
16:25 dogbert17 joined
Zoffix \o/ 16:32
timotimo since there's not only getpeername but also getsockname i'll be returning two ports and two ips when an async socket is accepted 16:33
Zoffix AlexDaniel: maybe you have a recommendation? I think it was you whose system had a ? that looked a lot different than \ twitter.com/kaokun/status/895682264998326273 16:35
AlexDaniel ?\ 16:36
yeah, silly me expected to look at it in my browser to see how it renders there
Zoffix Renders like this for me on this box: i.imgur.com/FbmsXq1.png 16:37
I don't think it's that much of a problem in real code, since \ is an unspace or a Capture, which tend to be in different places than ? 16:39
AlexDaniel Zoffix: my recommendation is really stupid. I'm using bitmap fonts for everything, these typically don't have the whole unicode range, so it falls back to any font that the system has with that glyph. This way, ascii range renders sharp and everything else is noticeably blurry
Zoffix my @stuff = keys @a ? @b; don't make much sense
*if it were \
AlexDaniel Zoffix: it works for me, but I can't see how anybody could benefit from that idea
Zoffix heh
I'll leave it unanswered then :) 16:40
AlexDaniel example: files.progarm.org/2017-08-10-19395..._scrot.png
it has slightly different angle too, so maybe the right solution would be to ping font authors to make it not look like \
I think Damian's answer would be to change your vim config to color unicode symbols differently 16:42
and the right answer is possibly to change your emacs config to change the background color without affecting syntax highlighting (?) :) 16:43
16:55 lizmat joined
AlexDaniel Zoffix: I don't know if it helps but I created this: github.com/perl6/doc/issues/1444 17:00
Zoffix You're chatting using emacs? 17:01
....all of a sudden the jokes about it being the whole operating system make sense :) 17:02
AlexDaniel yeah, I'm using ERC 17:03
timotimo AlercsDaniel 17:06
Geth MoarVM/asyncsocket_listen_sockname: 658b53456a | (Timo Paulssen)++ | src/io/asyncsocket.c
the listen callback shall also push socket and peer ip & port
 17:11
timotimo AFK before i see to it that the rakudo portion also works
AlexDaniel timotimo: while you're at it, you may also like RT #131766 :) 17:17
synopsebot6 Link: rt.perl.org/rt3/Public/Bug/Display...?id=131766
timotimo oh, hmm 17:18
that's more api design questions i suppose
but it's correct to put uv_tcp_bind before connect to get a source ip 17:21
for some reason i get empty strings here 17:26
no matter, gotta run
17:31 lizmat joined 17:51 lizmat joined
AlexDaniel .tell MasterDuke I've created a ticket. Please leave some comments there: RT #131879 17:51
yoleaux AlexDaniel: I'll pass your message to MasterDuke.
synopsebot6 Link: rt.perl.org/rt3/Public/Bug/Display...?id=131879
AlexDaniel FWIW I'm marking RT #131879 as a blocker, but I can easily change my mind. It is a very frustrating bug for whose who use rakudo to run long-running processes, but it could also be that I'm the only person affected. 17:54
synopsebot6 Link: rt.perl.org/rt3/Public/Bug/Display...?id=131879
17:56 lizmat joined
AlexDaniel for example, it only “leaks” with that specific gist, everything else seems to be normal 18:00
jnthn AlexDaniel: I'm going to look into a leak tomorrow that may well be the same one, fwiw 18:08
18:12 AlexDani` joined
AlexDani` jnthn: cool, thanks 18:14
japhb jnthn: Is there an active grant request that I can support? 18:15
Geth MoarVM/asyncsocket_listen_sockname: b943f8236d | (Timo Paulssen)++ | src/io/asyncsocket.c
you use VMString to decode strings, box it in BOOTStr
 18:32
MoarVM/asyncsocket_listen_sockname: fa64450d91 | (Timo Paulssen)++ | src/io/asyncsocket.c
return undefined hosts and ports in error cases
18:52 geekosaur joined
timotimo i wonder if i should do anything like this for the connect side, too? probably interesting to get the local host and port, but maybe not the target? 18:54
19:23 brrt joined
timotimo though having accessors for all four things would make things consistent 19:32
just set everything
i've got a moar bump coming up to make rakudo and moarvm agree on the signature of the callbacks for listen and connect 20:14
20:50 AlexDaniel joined 21:01 colomon joined 21:17 colomon joined 21:22 lizmat joined
samcv o/ 21:50
timotimo o/ 21:53
nwc10 o/ 21:54
oh, hangon, I should go to bed
timotimo good night nwc10
samcv: i only saw the first and about ½ of the second talk, but the parts i did watch were good 21:55
wow holy crap that last rt bug just now 21:56
Geth MoarVM/asyncsocket_listen_sockname: dd95f89b93 | (Timo Paulssen)++ | src/io/asyncsocket.c
also pass socket and peer names for outgoing connections
 22:03
timotimo (it didn't take me this long to do that, i just had the code uncommitted and unpushed for a while) 22:04
MasterDuke anyone have thought/comments on github.com/MoarVM/MoarVM/pull/626 (other than i need to rebase to fix the oplist conflict)? 22:09
22:12 dogbert17 joined
timotimo i say go ahead and merge it 22:20
MasterDuke rebased and force pushed 22:23
i don't have a commit bit, but maybe one of you fine people who do could merge once travis oks it 22:24
22:28 zakharyas joined
Geth MoarVM: 955a9cd1a8 | (Samantha McVey)++ | src/core/interp.c
Move indexim_s and eqatim_s to oplist order

Move them in interp.c to oplist order.		 22:32
samcv zhuoming should be happy the ops will be back in correct order 22:33
wonder if we want to make a script to put them all in order of their numbers? though the compilier may optimize it since it sees goto statements after each
jnthn japhb: Sorry, was AFK relaxing some. 6guts.wordpress.com/2017/05/12/loo...t-funding/ is the last thing I posted on the topic, and the offer there is still current, at least.
yoleaux 21:15Z <AlexDaniel> jnthn: FWIW some kind of breakage due to f590863e17 gist.github.com/Whateverable/2a908...7ba3339af2
AlexDaniel this is RT #131881 now I believe 22:34
synopsebot6 Link: rt.perl.org/rt3/Public/Bug/Display...?id=131881
jnthn AlexDaniel: Which repo is f590863e17 a commit in?
AlexDaniel jnthn: rakudo, can't bisect into moarvm yet
not with a bot at least
22:34 markmont joined
AlexDaniel Geth: ver github.com/rakudo/rakudo/commit/f590863e17 22:35
Geth AlexDaniel, version bump brought in these changes: github.com/perl6/nqp/compare/2017....2-gd40d6dc
jnthn lol, what, I managed to make a bug such that it makes gnome-terminal-server sigabrt, thus presumably uncovering a bug in that too? :D
geekosaur yep, unicode decoding fails poorly 22:37
jnthn AlexDaniel: Is there a way with the bot to try it with an env var set before spawning?
AlexDaniel actually, would be great to submit a bug report there as well
jnthn I'll try and recreate it tomorrow of course
Though the first thing I'll do is check if MVM_SPESH_DISABLE=1 and similar hide it 22:38
AlexDaniel jnthn: not via IRC, no, but I can do it
jnthn I'm presuming it's one of my clever opts being less than clever... 22:39
AlexDaniel jnthn: actually… I think all you need is just run it with MVM_SPESH_DISABLE=1 on head
jnthn AlexDaniel: Ah, yeah. :)
AlexDaniel if somebody has an idea on how should the command look like, please comment on github.com/perl6/whateverable/issu...-321693966 22:48
“commit: MVM_SPESH_DISABLE=1 HEAD say 42” ?
“commit: HEAD MVM_SPESH_DISABLE=1 say 42” ?
… both?
jnthn: seems to be failing with MVM_SPESH_DISABLE=1 too 22:51
jnthn !!
Oh wow
I wonder what I coulda touched that isn't spesh related... 22:52
timotimo AlexDaniel: i think both are fine, you can tell from the presence of = in there, right? 22:53
AlexDaniel timotimo: yeah, I think so. Should I support arbitrary env variables? 22:54
and how likely is it that someone's code is going to start with foo=smth ? :)
timotimo hmm
jnthn I figure so, though be careful people can't shell-inejct them
timotimo very bad idea, if you don't sandbox
imagine MVM_SPESH_LOG=/home/alexdaniel/.ssh/id_rsa 22:55
AlexDaniel not yet :'(
jnthn I mean, if you do allow any env vars, make absolutely sure that you parse them and set up a %*ENV
Or similar
timotimo even without shell injection to the shell vars you can remote trash files on the FS whereever the committable bot has write access 22:56
jnthn timotimo: Uh, but if we can do that by sending the spesh log there then a program can write there too already? :)
timotimo er, i guess
AlexDaniel I've been trying to find some time to implement sandboxing for a year… this is horrible
jnthn: not really. Well yes, sure, but not as easily, because of RESTRICTED setting 22:57
but of course working around it is very trivial
timotimo aye
AlexDaniel this is the reason why these bots don't respond to private messages. If somebody wants to destroy my server, please do it on public ?_? ?_? ?_? ?_? 22:59
.oO( this strategy is working surprisingly well )
 23:00
timotimo ah yes 23:02
23:23 MasterDuke joined
AlexDaniel what's the right falsy value for dev env vars? 0 or emtpy string? 23:38
both I guess github.com/MoarVM/MoarVM/blob/b27f...oar.c#L203 23:39
23:42 committable6 joined
AlexDaniel commit: MVM_SPESH_DISABLE=1 HEAD say %*ENV<MVM_SPESH_DISABLE> 23:42
committable6 AlexDaniel, ¦HEAD(ec7bc25): «1»
AlexDaniel commit: MVM_SPESH_DISABLE=1 f590863e17^,f590863e17 use lib ‘data/all-modules/moritz/json/lib/’; use JSON::Tiny; use Test; my @t = '{ "a" : "b\u00E5" }' => { 'a' => 'bå' }, '[ "\u2685" ]' => [ '?' ], qq{"x\c[ZERO WIDTH JOINER]a"} => "x\c[ZERO WIDTH JOINER]a", qq{"\c[ZERO WIDTH JOINER]"} => "\c[ZERO WIDTH JOINER]"; plan (+@t); for @t -> $p { my $got = from-json($p.key); is-deeply $got, $p.value; } 23:43
committable6 AlexDaniel, gist.github.com/2f9573fe562b78bcd1...546739f2bb
AlexDaniel :)
jnthn Hmmm 23:46
AlexDaniel++
timotimo m: -> Mu \a? { say a }() 23:54
camelia ===SORRY!=== Error while compiling <tmp>?Malformed parameter?at <tmp>:1?------> -> Mu \a?? { say a }()? expecting any of:? constraint?
timotimo m: -> Mu $a? is raw { say a }()
camelia ===SORRY!=== Error while compiling <tmp>?Undeclared routine:? a used at line 1??
timotimo m: -> Mu $a? is raw { say $a }()
camelia (Mu)?
timotimo maybe i'll implement passing hosts and ports with this so i don't have to immediately bump moar when moar learns to submit those 23:55
Zoffix AlexDaniel++
timotimo it could then also skip sending Str, Int, Str, Int in case of errors

Powered by Cro and the Raku® Programming Language.

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

Thank you!