github.com/moarvm/moarvm | IRC logs at colabti.org/irclogger/irclogger_logs/moarvm
Set by AlexDaniel on 12 June 2018.
Geth MoarVM: f06c695a30 | (Samantha McVey)++ | 2 files
[UNI] Make Block name lookup work for unassigned cp's

Unlike other properties which if they are unassigned go to default values, Block does not work this way. Instead of looking it up in the database, use similar code to how we check the block in other circumstances and use binary search over our Block info data structure.
... (6 more lines)		 00:52
MoarVM: e17a53b870 | (Samantha McVey)++ | docs/release_guide.md
Update the release guide a bit
brrt \o 09:23
nwc10 o/ 09:24
brrt \o 12:30
timotimo: you mean to say that the JITted version is slower than the non-jitted version? 12:32
brrt we can easily make a trunc_i8 though 12:37
brrt hmm, I may have to get sucked into improving the tiler... 13:01
bugger 13:03
well, the tiler table generator
timotimo brrt: no, the jitted version is faster, but the version without the bail is even fasterer 13:26
brrt ah, i see 13:45
I need to do a rethink on how we do tiler table generation in general, I think we can do that much better 13:46
timotimo that's a moarvm-compile-time-thing, yeah? 13:50
or does it impact the data structure we're using internally?
brrt not really, no 13:52
although I can replace my handwritten binary searches with a call to bsearch and be happier
timotimo at the moment multiplying by 4 is almost 2x faster than left-shifting by 2 :) 14:34
we just don't have an implementation for the shl and shr ops in the jit, though 14:35
jnthn
.oO( no shift, sherlock... )
 14:37
timotimo this particular benchmark is faster with Int than with int 14:38
shameful ;)
jnthn Yeah...improving the performance of code using native types is near the top of my post-vacation todo list :) 14:41
Gotta finish/give my workshop talks before I get vacation though :P
timotimo of course 14:43
:)
jnthn At least I've got a nice view while I'm writing :) 14:44
timotimo grmbl, this program gets faster if i replace the int8 arrays with regular arrays that store Int objects. though perhaps that has something to do with only using 0 and 1? 14:56
this program simulates 500 steps of an 80x80 2d cellular automaton that takes four neighbors + itself as input and has a probabilistic input in 11.25 seconds, which includes displaying it with unicode HALF BLOCK characters and reverse video, which runs in parallel to the simulation 14:59
19.53 seconds user, 177% cpu usage
that's not even bad.
i'm using an array of arrays, though, not a 2d shaped array
nine github.com/niner/Inline-Perl5/issues/122 is....hard 19:25
The bug is quite clear: arg_info only contains information about MAX_ARGS_FOR_OPT arguments but the getarg_i optimization/fix needs information about the corresponding argument, regardless of its number. 19:26
I guess that's what I get for implementing a fix as an optimization. An optimizer can just have arbitrary limits for cases it tries to optimize while the fix always needs to work. 19:27
nine Actually there's another bug lurking there. A getarg_i may actually be in a different basic block than the corresponding arg_i instruction. In that case arg_info may not be initialized at all. 19:30
MasterDuke has anyone else seen this? tratt.net/laurie/blog/entries/why_a...art_1.html 19:37
interesting article about language VMs and how/when they reach steady state behavior (i.e., initial interpreting -> optimizing -> running optimized code) 19:39
robertle watch 19:45
the y axis on the graphs though!
MasterDuke yeah, the HN commentary has some good points 19:47
nwc10 MasterDuke: no I hadn't. It's awesome. Not that I know how to imrpove stuff based on it 20:04
nwc10 but the "people weren't really sure so we ran it 10 *times* longer" (rougly) approach, and then what they found, is jaw dropping 20:05
and even the graph for "C does strange things" - because they thought to compare all the VMs with somethign that should be predictable 20:06
it all seems very thorough
Kaiepi for the async sockets fd pullreq, would it be a good idea to add an op that just initializes the socket, then have listen/connect use the socket it creates instead of initializing it on listen/connect? 20:06
i can't think of any other way to have the socket immediately available for IO::Socket::Async.listen to use 20:07
what i'm thinking is having an op like nqp::asyncsocket that returns a raw socket, which could then be passed to nqp::asyncconnect and nqp::asynclisten, but that's a breaking change 20:14
and i'm not sure how you'd get the socket's struct to get cleaned up after freeing the structs listen/connect use 20:15
Geth MoarVM: 2fb3265a18 | (Stefan Seifert)++ | src/spesh/optimize.c
Fix segfault of native subs with more than 8 arguments

arg_info only contains information about MAX_ARGS_FOR_OPT arguments but the getarg_i optimization/fix needs information about the corresponding argument, regardless of its number. Instead of relying on other optimizations, just look backwards for the corresponding arg_i instruction.
Fixes github.com/niner/Inline-Perl5/issues/122		 20:23
[Coke] reading MasterDuke's page there; I've seen jnthn measure actual CPU ops rather than trying to lock down all variables; seems like that's a much easier data point to get and have a high confidence level on. 20:24
nine Not perfect, but this'll do for now as I'm heading to bed and will be AFK all of tomorrow.

Powered by Cro and the Raku® Programming Language.

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

Thank you!