Perl 6 language and compiler development | Logs at colabti.org/irclogger/irclogger_log/perl6-dev | For toolchain/installation stuff see #perl6-toolchain | For MoarVM see #moarvm
Set by Zoffix on 27 July 2018.
00:27 fake_space_whale joined 01:00 lizmat joined 03:24 fake_space_whale left 03:33 vrurg left 03:53 vrurg joined 05:44 vrurg left 06:30 lizmat left 06:52 ufobat_ joined 06:56 ufobat__ left 07:54 [TuxCM] left 08:04 lizmat joined 08:14 [TuxCM] joined 08:19 lizmat left 08:26 lizmat joined
nine 19:08 < sangy> it's crazy how hard it is to keep things reproducible 09:05
well put
09:24 ufobat_ left
lizmat # expr bail: Cannot handle DEOPT_ONE (ins=sp_guardconc) 09:56
fg 09:57
[TuxCM] bg 10:07
timotimo mfg 10:09
[TuxCM] Rakudo version 2019.03.1-245-g07b4f7438 - MoarVM version 2019.03-101-g8eb232dbc
csv-ip5xs0.683 - 0.699
csv-ip5xs-205.872 - 5.946
csv-parser21.784 - 22.855
csv-test-xs-200.429 - 0.432
test6.790 - 7.602
test-t1.660 - 1.667
test-t --race0.772 - 0.775
test-t-2027.501 - 27.858
test-t-20 --race8.763 - 8.862
lizmat just pushed github.com/rakudo/rakudo/commit/84...6fea9217f0 10:25
First attempt at providing an interface to spesh logs
timotimo oh! 10:26
lizmat: have you seen moarvm's tools/ folder?
it's not quite relevant to what you're working on
but there's the graph_spesh tool that turns a piece of spesh log into a diagram for graphviz 10:27
lizmat it gives some inspiration :-)
timotimo \o/
nine Apparently the remaining reproducibility issues are all because of different order of register usage in the generated bytecode. The really weird thing is that this only appears in Perl 6 code. NQP code is perfectly reproducible despite using the same compiler backend with the same register allocator 10:33
Where's the code generator for creating Scalar containers for arguments? 11:02
jnthn nine: Probably that's to do with lexical to local lowering, in Perl6::Optimizer 11:51
I dobut it goes over the lexicals to lower in anything other than hash order 11:52
*doubt
And so would ask for the locals in different orders 11:53
nine jnthn: I've already made lexical_vars_to_locals iterate on sorted hash keys 12:04
12:16 AlexDaniel left 12:17 AlexDaniel joined
jnthn OK, not sure then... Parameter lowering is in Perl6::Actions, lower_signature or similar, but I think that just works in parameter order, so can't imagine what might be random there 12:39
12:45 MasterDuke joined 12:46 MasterDuke left, MasterDuke joined
nine At least I could now confirm that hash randomization is indeed responsible for all remaining differences 12:56
What's still confusing is that I don't see any differences in the order of the generated code. Just in the usage of registers by the code. So somewhere the order of processing does not actually change the order of the generated code but only the state of the register allocator. 12:57
Since with the new MAST backend code is generated pretty much in order of walking the QAST, that shouldn't leave many possibilities 12:58
13:00 [TuxCM] left
timotimo fascinating 13:00
13:00 [TuxCM] joined
MasterDuke nine: so the register allocator needs a sort thrown in somewhere? or a way to seed the hash randomization? 13:02
nine MasterDuke: the register allocator itself works completely predictable. It uses only arrays and is less than 40 LOC. And it's also used for NQP where we don't have any reproducibility issues anymore 13:06
Apparently the optimizer is at least innocent 13:11
AlexDaniel github.com/perl6/problem-solving/i...-487286532 13:48
that application is really strong, in my opinion 13:49
should we wait for somebody else to try also?
nine I'm officially running out of ideas 13:50
13:55 lucasb joined 14:10 leont joined 14:16 rba joined 14:20 leont left
nine Oooooooh github.com/perl6/nqp/blob/master/s....nqp#L1346 14:37
With that fixed, the moar --dump results of 2 builds are identical! But there's still differences in the bytecode. Apparently lists of 16 bit numbers are in different order. 14:39
gfldex m: my %h = 1 => 2, 3 => 4; say %h.head(%h.elems / 2); 14:54
evalable6 (exit code 1) This type cannot unbox to a native integer: P6opaque, Rat
in block <unit> at /tmp/Bwk9Fuhm2T line 1
14:55 patrickb joined
lizmat m: <a b c>.head(1/2) # golfed 14:57
evalable6 (exit code 1) This type cannot unbox to a native integer: P6opaque, Rat
in block <unit> at /tmp/7SwkT_30q8 line 1
gfldex I'm undecided if that is just LTA or if asking for .head(Rat $n) in itself is wrong. 15:02
lizmat it is an LTA error, it should coerce to Int 15:03
gfldex: would appreciate a ticket
fixing it now
gfldex lizmat: will do
lizmat: #2862 15:05
lizmat gfldex++
15:13 [TuxCM] left
lizmat misses the bots 15:31
15:35 [TuxCM] joined 15:46 leont joined
AlexDaniel lizmat: uh, which ones? 16:08
lizmat Geth
AlexDaniel Geth: help
Geth AlexDaniel, Source at github.com/perl6/geth To add repo, add an 'application/json' webhook on GitHub pointing it to hack.p6c.org:8888/?chan=#perl6 and choose 'Send me everything' for events to send | use `ver URL to commit` to fetch version bump changes
AlexDaniel so it's here but it doesn't work?
lizmat so why is Geth not reporting my commits to rakudo ?
lizmat did 2 in the past 30 mins 16:09
AlexDaniel ahhhh… something about dns or whatever
it's probably running on another server now?
16:13 [TuxCM] left
MasterDuke lizmat: github.com/rakudo/rakudo/blob/mast...sh.pm6#L33 will miss `JIT: bailed completely because of <foo>`, right? 16:15
lizmat MasterDuke: good catch, will fix 16:16
bartolin_ I'd like to suggest to unify nqp::substr across the different backends. Should I open an issue in the problem-solving repo for that or would a plain nqp issue suffice? (Nothing should be changed for Perl 6, only for NQP.) 16:17
lizmat MasterDuke: eh... maybe not needed, as the it will catch the next line ? 16:18
bartolin_: I don't think a problem-solving issue is needed
backend differences should be as small as possible :-) 16:19
bartolin_ ok, thanks. I'll write down some context, options, etc. for an NQP issue 16:20
MasterDuke lizmat: if a spesh log only has a `JIT: bailed completely because of <foo>`, but not a `expr bail: Cannot get template for: foo`, it won't get picked up 16:31
lizmat can that happen ? 16:32
MasterDuke don't know
lizmat haven't seen it happen in the spesh logs that I've seen so far 16:33
MasterDuke but those are two different kinds of bails. bailed completely means it isn't getting jitted at all, expr bail just means there's no template. might be nice to be able to differentiate
lizmat ah, ok... hmmm... 16:34
16:35 [TuxCM] joined
MasterDuke oh, and we have in the not-too-distant past had the situation where there were templates for ops, but it wasn't lego jitted. i don't remember how that looked in the log, and that isn't really a usual situation, but it might cause a 'bailed completely' without an 'expr bail' 16:42
ugexe github.com/MoarVM/MoarVM/pull/1098 <-- these have expr templates but no legojit 16:44
16:48 raschipi joined
lizmat MasterDuke: re cuids in spesh log 16:48
16:48 [TuxCM] left
lizmat I was living under the impression that each cuid would refer to the same frame 16:49
MasterDuke ugexe++, those do create `bailed completely` in the spesh log, but *not* `expr bail`
lizmat but that does not appear to be the case
16:49 [TuxCM] joined
lizmat so what is the function of cuid? 16:49
lizmat was hoping to be able to use it as some "primary key" for selection 16:50
MasterDuke lizmat: i think i thought that too recently, but i don't remember exactly if/how that was wrong
jnthn, timotimo, or brrt would know
hm, "20:58 timotimo but yeah, having a sub specifically for the op in question gives us a different cuid, which will have its own set of spesh candidates and everything" is the only thing slightly relevant i see in the channel logs 16:52
it looks like cuids by themselves might not be unique, but cuild + file is 16:57
90% sure 16:59
lizmat yeah, came to that conclusion as well 17:00
which sorta makes the cuid superfluous
for spesh log introspection purposes anyway
17:22 leont left 17:23 lizmat left
MasterDuke why aren't roles augmentable? 17:45
17:47 leont joined 18:01 Kaypie left
patrickb o/ 18:12
raschipi \o
patrickb I'm pondering Windows process exec argument quoting behavior. There's already a bug report about it (github.com/rakudo/rakudo/issues/2005) and I did a small writeup about my research (gist.github.com/patzim/73d4e9c7f0c...ssing-md). 18:15
This starts to feel rather spec-y.
Is this a candidate for perl6/problem-solving?
18:24 leont left
jnthn .tell lizmat cuid is a unique ID within the compilation unit, which can be identified with the filename; mostly it's useful if the log has multiple routines of the same name from the same file in it 18:35
yoleaux jnthn: I'll pass your message to lizmat.
jnthn .tell lizmat to be clear, it's an ID for a Frame (which might be a routine, or block, or thunk) 18:37
yoleaux jnthn: I'll pass your message to lizmat.
jnthn MasterDuke: Partly because composition is flattening, so we copy out of the role, and partly because of the genericity handling (it matters that the methods are in a single body block) 18:38
MasterDuke: Classes are not generic for the same (well, opposite) reason :)
19:25 lucasb left
MasterDuke jnthn: any reason you couldn't do it if you were ok with anything that *previously* did the role *not* getting whatever augment you did? 20:11
20:28 ufobat joined
jnthn I still don't really know how you'd get around the fact that the handling of genericity depends on closure semantics. 20:39
It falls out of the "lexpads are closed by compile time" thing
20:52 squashable6 left 20:57 squashable6 joined
MasterDuke jnthn: i don't really understand what you mean by handling of genericity? i feel this is something different than roles being parammeterizable 21:07
jnthn 2 words for the same thing really
But in role Foo[::T] { } then inside of there ::?CLASS and T are both lexicals bound by the signature 21:08
So it's like saying "if I have a sub foo() { my $a }, how can I put another sub inside of foo in the future", and the answer is you can't 21:09
(The point being, the augmented methods would have to go inside of the role body to close over the correct arguments) 21:10
MasterDuke hm. but isn't that what augment enables? e.g., put another method inside a class?
jnthn Not really
Not in a lexical sense
m: class C { my $a = 42 }; augment class C { say $a } 21:11
evalable6 (exit code 1) 04===SORRY!04=== Error while compiling /tmp/AU9pxI14fE
augment…
jnthn, Full output: gist.github.com/dc9c81a6c0275dcf91...4a25acc896
jnthn m: use MONKEY; class C { my $a = 42 }; augment class C { say $a }
evalable6 (exit code 1) 04===SORRY!04=== Error while compiling /tmp/22eR58HDrL
Variable '$a' is not declared
at /tmp/22eR58HDrL:1
------> 03C { my $a = 42 }; augment class C { say 08⏏04$a }
MasterDuke ah, so we sort of could allow augmenting a role if it never used T or ::?CLASS or pre-existing lexicals? 21:12
but since that's such a significant reduction in usability why bother?
jnthn It's zero usability, because the type constraint of self of every method in the role is ::?CLASS 21:13
So there's literally nothing you can augment in that doesn't require it
MasterDuke oh, right
jnthn Well, I guess in theory if you wrote method foo(Mu: $bar) { } you sorta could 21:14
But that won't end too well if it's a multi and competing with other candidates :) 21:15
AlexDaniel patrickb: yes, especially if you want to end up with a document in the end 22:16
22:26 patrickb left 23:31 leont 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!