nwc10 good *, #moarvm 08:37
jnthn o/ 08:41
timotimo #moarvm, * good 08:45
FROGGS__ o|) --> (·) 10:36
nwc10 jnthn: I baked you another patch, at [email@hidden.address] 11:53
1% reduction in startup CPU count
(as "threatened" some weeks back) 11:54
jnthn nwc10: I get permission denied...is there a ro url? 12:14
(I just did git remote add -f nwc10 [the url]) 12:15
nwc10 gitlab.com/nwc10/MoarVM.git
I hope
jnthn yup, I can git fetch nwc10 now :) 12:17
nwc10 excellent
jnthn Cute patch 12:18
nwc10 I can't find anything else similar that's as easy
I can benchmark it at (seemigly) 1% less time to run "Hello world" but I don't trust benchmarks that are under 5% :-/ 12:19
dalek arVM: af7753e | nicholas++ | src/6model/serialization.c:
Inline assert_can_read(), which is called 7,133,540 times for perl6-m -e 1

assert_can_read() is a very small function, but 15th highest by Ir count in callgrind's output for perl6-m -e 1. Hence the overhead of making the function call is likely to be similar in size to the function's body.
cachegrind reports that inlining it reduces I refs from 515,809,288 to 509,668,666, ie a 1% reduction in CPU instructions needed to run perl6-m -e 1		 12:21
jnthn nwc10: Measuring instructions retired in pretty reliable, though, at least :) 12:22
dalek arVM: c33c704 | (Jimmy Zhuo)++ | src/6model/reprs/P6opaque.c:
removed an repeated function
 13:15
arVM: 364432d | (Jimmy Zhuo)++ | src/6model/reprs/P6opaque. (2 files):
Inline some small functions
 13:28
psch m: say "m\x0308" 17:23
camelia rakudo-moar 4c1f57: OUTPUT«m̈␤»
psch eh
i'm not used to tab 4 not being camelia :P 17:24
dalek arVM: bbf73a1 | FROGGS++ | src/core/nativecall.c:
decont arg in not 'is rw' param slot

This shall unbreak Inline::Perl5. nine++ for reporting it.		 17:51
arnsholt Has anyone considered running afl-fuzz on Moar? 18:22
jnthn It's been mentioned here before, but nobody had les tuits to make it happen yet. 18:24
arnsholt Who's discussed it? I have a desktop machine at work that's standing mostly idle. Figured it might be worthwhile to use a core for fuzzing Moar 18:25
But might be useful to discuss it a bit with someone before just blindly setting of in some direction
jnthn I can't remember, 'fraid 18:28
I think if you search the irclogs for this channel you may dig something up
arnsholt Yeah. I'll grep the clogs
Thanks!
FROGGS m: my uint8 $a = 200; say $a # is autoboxing to blame? because P6int->get_int is called with bits=64 and unsigned=0 only...
camelia rakudo-moar c314f4: OUTPUT«-56␤»
jnthn I know I saw afl before, and I'm pretty sure it was on this channel :) 18:29
FROGGS would like to know where autoboxing happens 18:29
jnthn FROGGS: We totally ignore the u there at present
FROGGS jnthn: well, P6int->compose gets it and it gets serialized... but it gets lost later on
jnthn Yeah, but on lexicals I meant. 18:30
FROGGS is that something I could fix?
jnthn But the answer is likely that QSAT to MAST should be emitting a box_u not a box_i
FROGGS aha 18:31
jnthn I can probably delegate implementation of that stuff to you, but I need to ponder on design a little more.
FROGGS k
jnthn The Moar opset already has a bunch of _u though
We just ain't doing anything with 'em yet. 18:32
FROGGS nods
jnthn I also dunno how we're gonna handle this stuff on JVM :S
FROGGS "differently" :o) 18:33
nwc10 jnthn: I baked you another patch at gitlab.com/nwc10/MoarVM 18:37
only 0.5% less CPU this time
jnthn Thanks :) 18:41
arnsholt I think on JVM you need to do the interpretation of unsigneds on your own 18:42
But IIRC the VM spec mandates that integer types are stored twos complement, so there's no ambiguity there
jnthn masak: I'm inclined to reject RT#124226. That's stretching "syntactically" incredibly far.
masak: We also don't keep enough information around to do it realistically cross-compilation unit and I'm not sure we should. 18:43
masak: Finally, we do our static analysis on lexical things, not on method calls. We'd have to factor in the meta-object's behavior too, which we don't really know. 18:44
uh, shoulda been on #perl6...d'oh 18:45
dalek arVM: f37765f | nicholas++ | src/6model/reprs/MVMHash.c:
Inline extract_key(), which is called 57306153 times by at_key() and bind_key()

at_key() and bind_key() are 2nd and 3rd highest by Ir count in callgrind's output for perl6-m -e 1. Both call extract_key(), which is a simple wrapper around the macro MVM_HASH_EXTRACT_KEY(). Making the function inline reduces the CPU instructions needed to run perl-m -e 1 by about 0.5%		 18:45
nwc10 at this rate, (obviously) I just need to find about infinity more things and we'll be another 0.5% faster 18:50
dalek arVM: 0acf845 | FROGGS++ | src/strings/utf8.c:
fix throwing ex when utf8 string is invalid, fixes RT #124185
 19:05
japhb nwc10: As you have only two entries in the sequence, there's not enough evidence to assume exponentially decreasing returns, as opposed to (say) a normal distribution around .75%. :-) 19:12
I suppose that we could assume the distribution is bounded by 100% (because causality) and 0% (because you wouldn't apply a known pessimization) 19:13
arnsholt Heh. The afl-fuzz Makefile is funny: "NOTE: If you can read this, your terminal probably uses white background. This will make the UI hard to read. See docs/status_screen.txt for advice." 20:12
FROGGS hehe
japhb Nice
I once tried to write a colored-output script to autodetect whether the terminal had a light background or dark one and compensate, and I discovered that this seemingly easy thing is essentially completely brain-dead and non-portable. 20:13
arnsholt Yeah. The afl docs just go "use white-on-black or turn colours off entirely" 20:14

Powered by Cro and the Raku® Programming Language.

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

Thank you!