01:44 tokuhiro_ joined 02:16 tokuhiro_ joined 02:23 FROGGS_ joined 02:51 Util joined
dalek arVM: b3d0345 | hoelzro++ | src/ (4 files):
Restore "Track whether or not a capture owns its callsite"

This reverts commit 9befd69353643bff3a529aa5bea8101cd574fa29.		 03:31
arVM: 132f390 | hoelzro++ | src/core/args.c:
Just check arg_flags to know whether or not to get rid of them

One, arg_flags should know best. Two, checking the callsite object is checking an object we don't own, and may be invalid by this point in the program.
03:46 vendethiel joined 06:23 TimToady joined 06:56 domidumont joined 07:01 domidumont joined 07:08 FROGGS_ joined 07:34 TimToady joined
nwc10 hoelzro++ # ASAN approves 07:50
(UTF-16)-- # with the best "features" of UTF-8, the best "features" of UCS-32, some special features of its own - no wonder ASAN hates it :-) 07:51
08:49 zakharyas joined 09:01 kjs_ joined 09:37 kjs_ joined 11:31 brrt joined
nwc10 hi brrt 11:45
I found some things. hhvm.com/blog/10205/llvm-code-generation-in-hhvm
... Ultimately, we decided to not deploy the LLVM backend to production for now. Making such a large change across the whole web fleet always comes with risk, and without a performance improvement to justify the risk it’s not worth it.
but also, this one made me wonder "oh geez, is this crazy or inspired?" -- Modifying code that may be running in another thread is a core part of HHVM’s design. 11:46
meanwhile, I noticed that on www.zend.com/en/resources/php7_infographic they are happyly showing (er, choosing) graphs that show PHP7 to be faster than HHVM
HHVM 3.7.0
current HHVM is 3.10.10 and between the two we had hhvm.com/blog/9293/lockdown-results...erformance 11:47
-- During lockdown we achieved a 19.4% RPS improvement for MediaWiki workloads, and a 1.8% RPS improvement for WordPress. We demonstrated that HHVM is 55.5% faster than PHP 7 on a MediaWiki workload, 18.7% faster on a WordPress workload, and 10.2% faster on a Drupal 7 workload.
so I wonder how fast PHP7 is compared with "this week"'s HHVM
and if Zend are stuffed
(well, unless you want to run PHP on non x86_64, because, for example, ARM is interesting) 11:48
11:48 ziggypup joined 11:49 ziggypup left
brrt hi nwc10 12:13
(i was off for lunch)
nwc10 ilmari: ^^ :-) 12:14
hi brrt
brrt is reading about that page 12:15
i recall rurban saying good things about php7
at least with regards to speed
brrt is thinking of switching blogging platforms, but then recalls that he doesn't care about managing websites 12:22
timotimo jnthn: i think the only thing i'd need to change to make ordat not leak synthetics is use MVM_string_ord_basechar_at instead of MVM_string_get_grapheme_at - nqp::ordfirst also uses get_grapheme_at and thus probably wants to be changed as well. 12:25
brrt nwc10: very interesting. they're doing pretty much everything like how we're doing the moar-jit 12:31
i do have to ask
do my blog post come of as boasting in the same way this one does?
i mean, they almost *want* to make it sound difficult and complex and Serious Engineering 12:32
timotimo i don't think your blog posts sound overly boasting 12:33
brrt oh, right, this is facebook, that .. explains? :-P 12:34
timotimo huh? what is facebook? 12:35
oh, the hhvm people
nwc10 brrt: I believe that a team of them have been workign on HHVM for several years
and I got the impression that facebook tries really hard to only hire smart folks 12:36
so they probably think that it's hard because they are finding it hard
but also, they do need to make legacy code go fast
rather than (re)designing the language if something turns out to be hard/impossible to make fast 12:37
brrt hmmm.... 12:38
yeah, i think that the whole problem of making (legacy) dynamic language code go fast is slowly proving more difficult than anybody imagined 12:39
fwiw, i'm currently using pypy for Real Work, and it is a good 5 times faster on my (evil O(n^2)) algorithm 12:40
i'm stuck with python because $people
nwc10 presumably $people are on 2.7 - do they have a plan to migrate?
brrt my personal plan is to write everything py3 compatible
timotimo pypy is fantastic. i'm willing to say that over and over again :)
brrt python is a pretty good language at any rate :-) 12:41
not to say i don't miss perl, but the meanest thing i can think of is that python is slow; that must be pretty high praise 12:42
nwc10 I continue to assume that a big problem that Python 3 has with getting adopted is that Python 2.7 was already too damn awesome
timotimo oh, huh
i totally overlooked the "get basechar info" part to ordfirst and ordat 12:43
brrt anybody happens to be familiar with geometric algorithms and knows a less-than-O(n^2) to find all line intersections between two polygons?
timotimo nqp-m: say(nqp::ordfirst("à")) 12:44
camelia nqp-moarvm: OUTPUT«224␤»
timotimo nqp-m: say(nqp::ordfirst("ǎ"))
camelia nqp-moarvm: OUTPUT«462␤»
timotimo can has synthetic ...
nqp-m: say(nqp::ordfirst("\r\n"))
camelia nqp-moarvm: OUTPUT«13␤»
timotimo nqp-m: say(nqp::ordat("\r\n", 0))
camelia nqp-moarvm: OUTPUT«13␤»
timotimo nqp-m: say(nqp::ordat("\r\n", 1))
camelia nqp-moarvm: OUTPUT«Invalid string index: max 0, got 1␤ at /tmp/SxL62KDI03:1 (<ephemeral file>:<mainline>:22)␤ from gen/moar/stage2/NQPHLL.nqp:1303 (/home/camelia/rakudo-m-inst-2/share/nqp/lib/NQPHLL.moarvm:eval:190)␤ from gen/moar/stage2/NQPHLL.nqp:1506 (/home/camelia/rak…»
timotimo hum.
i still can't quite grasp why azawawi got trouble with JSON::Fast :( 12:50
brrt did.. he try turning the JIT off? 12:52
timotimo oh, you think the jit doesn't handle that? 12:56
it does, though 12:57
well, perhaps it doesn't handle it correctly?
signedness and such?
brrt handle what exactly :-) 12:58
timotimo if the number is negative, it calls nfg_get_grapheme_info
and noms the ->base
it just cmp and jge
brrt i dunno.. there may have been things i've been missing 12:59
in recent moar news
timotimo yeah, brrt, getting the jit involved breaks it 13:05
now you fix it! :P 13:06
brrt ok, so what is the issue exactly :-) 13:07
gimme a testcase
and i will move the world
timotimo perl6 -e 'use nqp; for ^200 { say nqp::ordfirst("\r\n") }'
starts out outputting 13, then outputs 4294967309
lines 1518 and following in our .dasc file are what does that 13:08
so the comparison against 0 must be wrong-ish
jnthn Ah, so the JIT is where that synthetic bug comes from 13:12
timotimo thank brrt for pointing it out
brrt aha.... 13:14
13:18 domidumont joined
timotimo i don't know much about signedness and unsignedness at the ASM level 13:18
13:33 tokuhiro_ joined
brrt timotimo; that isn't enough to replicate it for me? 14:17
timotimo it isn't?
it's 100% reliable on my end 14:18
with a higher number perhaps?
a funny thing: we don't jit the NFA stuff because nqp::getstderr - which it uses for debugging - isn't implemented in the jit 14:22
brrt wow.... 14:23
no, it's apparantly related to output
i had no idea we had regressed so far.. :-(
timotimo hm?
oh, i mean the code that builds NFAs
and the NFA optimizer 14:24
what do you mean is related to output?
anyway, i'm going to implement the bitwise ops that arnsholt's sum module uses 14:26
i have a bail log for his stuff
3 BAIL: op <bor_I>
3 BAIL: op <blshift_I>
2 BAIL: op <bxor_I>
2 BAIL: op <brshift_I>
2 BAIL: op <band_I>
1 BAIL: op <isprime_I>
1 BAIL: op <bnot_I>
(pasted here for my convenience :P) 14:27
14:32 kjs_ joined
timotimo brrt: you really can't reproduce it? there isn't an architecture difference, right? 14:35
brrt yeah, i can reproduce it
these are all bigint thingies, aren't they?
timotimo that's right 14:36
the only reason those aren't done yet is that there's an allocation that happens in interp.c that needs to be moved into bigintops.c 14:37
brrt seems like a plan, if that is safe 14:44
jnthn Should be; generally I like to keep op bodies in interp.c slim 14:45
timotimo getstd* all have a check for interp's handles to be null or not 14:46
i'm not sure if there's a good reason to duplicate that in the jit
brrt near as i can tell, the actual code fragment for MVM_nfg_get_synthetic in the jit is quite safe...
maybe i should follow it a little further
timotimo on the other hand, you can compile a frame that only actually runs these ops after a bunch of runs
and then it can segv :\ 14:47
oh well. it's not that hard to do an exception in the ji
jit*
brrt what do you mean timotimo? get a segv? 14:50
timotimo if tc->instance->std*_handle is null, then things go bad 14:51
jnthn Those are set up at startup, and I don't think they ever go away. 14:52
brrt i may be misunderstanding, but it seems that the issue *may* lie in p6box_i 14:55
but that would be.. weird
timotimo brrt: could it be we have to set the upper 32bit of teh number to zeroes? 14:56
brrt: because MVMGrapheme32?
brrt whatddayamean 14:58
yes, that could be
you are quite correct 14:59
timotimo i can hardly believe i caught that...
dalek arVM: 59be148 | brrt++ | src/jit/emit_x64.dasc:
Fix ordfist/ordat by correct size of synthetic

  timotimo++ for spotting and suggesting		 15:05
timotimo neato.
brrt you don't really have to explicitly zero the upper bits, x86 does that by itself
timotimo just have to use the right register size 15:06
brrt but it's stil very important not to load a bigger field than you have
aye
timotimo right :)
15:14 kjs_ joined
nwc10 brrt: I guess I'm repeating myself, but that's also what a "conclusion" is :-) 15:14
it's nice (for us) that two different competant teams, in two different languages, have investigaged using LLVM as a JIT (to the point of working code) and both found that it's LTA compared with what initial benchmarks would have though 15:15
thought
I think, sufficient to demonstrate (twice) that it's not cost effective.
brrt ... yes, that is very true
on the other hand, julialang.org/ 15:16
although i haven't looked at that in anything resembling detail
but, retrofitting a dynamic language to run on llvm, that isn't a good way forward, and i can imagine why 15:17
nwc10 I really don't know enough about Julia to know if it fits LLVM's worldview better than the "P" lnguages
brrt (for another data point, unladen swallow)
nwc10 oh, right, Julia was targetting LLVM from the start?
brrt yes
nwc10 Unladen Swallow didn't *seem* to get far enough that I'm confident that I'd trust it massively.
Also, that was some years ago now, and LLVM got better
so 2.5 :-)
mmm, as I understand it Scala makes compramises in language design to fit with what the JVM can do. Maybe Julia is similar 15:18
15:52 domidumont1 joined 15:54 kjs_ joined
brrt julia is a bit more primitive i think than some other languages 15:55
but the main competition for julia is, i think, not python or perl, but R
nwc10 ah :-)
and that all makes sense
brrt another awesome language by the way, which interestingly enough doesn't have the identity crisis of any of the P languages 15:56
15:57 domidumont joined
dalek arVM: 45d5efc | timotimo++ | src/ (4 files):
jit bigint ops or, and, xor, brshift, blshift and bnot.
 16:02
jnthn m: my @a of int; say @a; 16:03
camelia rakudo-moar cca5f2: OUTPUT«[]␤»
jnthn m: my @a of int; say @a.WHAT;
camelia rakudo-moar cca5f2: OUTPUT«(Array[int])␤»
jnthn yay, another thing fixed (locally)
16:19 domidumont joined
dalek arVM: 308bfa4 | timotimo++ | src/ (3 files):
jitting hintfor allows rakudo's "compile_var" to be jitted to completion.
 16:26
16:34 domidumont joined
japhb test 16:37
Huh
It's like my previous message just went *poof*
timotimo success
i didn't see any messages from you :(
japhb jnthn: Did you decide whether you would be in Stockholm first weekend in December and up for a meetup? (I definitely will; I'm flying in Friday evening.)
TimToady japhb: did it start with a slash? 16:38
japhb Nope; to resend I just up-arrowed and hit enter. It was just ... gone.
japhb wonders if it was software, or the scroll wheel on the mouse just happening to glitch at the same moment I tried to send ...
nwc10 timotimo: your most recent change breaks the NQP build :-( 16:40
timotimo oh? damn!
hintfor you mean?
damn, it does!
nwc10 This representation (P6int) does not support associative access at gen/moar/stage1/nqpmo.nqp:744 (gen/moar/stage1/nqpmo.moarvm:add_method:11)
...
timotimo ah 16:41
return type: void
that's wrong :)
jnthn timotimo: Please do always at least "make clean; make test" in NQP on Moar changes 16:42
timotimo right. i ran spec tests instead m)
jnthn Yeah, builds are long-running and stress spesh/JIT more, though :)
dalek arVM: e4e412f | timotimo++ | src/jit/graph.c:
fix return type of hintfor

and with it, the nqp build
timotimo there we go
nwc10 run all the things?
they're quite fast if you're not using ASAN 16:43
16:44 travis-ci joined
travis-ci MoarVM build failed. Timo Paulssen 'jitting hintfor allows rakudo's "compile_var" to be jitted to completion.' 16:44
travis-ci.org/MoarVM/MoarVM/builds/91852366 github.com/MoarVM/MoarVM/compare/4...8bfa47590f
16:44 travis-ci left
nwc10 the Rakudo setting build also stresses stuff a lot 16:51
but yes, the spectests find other things
hoelzro o/ #moarvm 17:08
nwc10 \o hoelzro 17:10
dear travis-ci, it's fixed.
hoelzro o/ nwc10 17:11
timotimo o/ 17:20
17:27 domidumont joined
hoelzro o/ timotimo 17:30
17:36 domidumont joined
timotimo uh oh 17:37
my latest patches seem to cause a "double free or corruption" in the original un-optimized sha256 code 17:38
nwc10 that's glibc malloc? If so, if you run it with valgrind, do you get more useful info?
timotimo i shall try. 17:42
i would have tried asan
nwc10 both are useful, but if you've already got a build and one testcase to run, valgrind gets you answers more quickly 17:43
(and finds generally more stuff)
timotimo the rate of reproduction for this crash is about 50% 17:51
it doesn't help that the whole compilation of the digest lib goes through valgrind now, too ... 17:53
ah, excellent! problem caught on tape 17:56
the thing was free'd inside jitted code somehow, probably inside the big integer ops i built? 17:57
huh, gc_free of MVMString.c of all places
tries to re-free it
er 18:00
it does an invalid free
if it tried to do a double-free, the message would surely have looked different 18:01
18:06 FROGGS joined
FROGGS o/ 18:10
timotimo ohai FROGGS
hooray! the problem exists before my patches were applied! 18:25
that makes it a bit harder to find the culprit, though :\ 18:34
FROGGS m: use NativeCall; class A is repr<CStruct> { has int8 $.a }; class B is repr<CStruct> { has int8 $.b }; class AB is repr<CStruct> { HAS A $.a; HAS B $.b }; say nativesizeof(AB) 19:31
camelia rakudo-moar 73485c: OUTPUT«9␤»
FROGGS I've got a patch for that
jnthn 9?! :) 19:32
FROGGS++ :)
dalek arVM: 7b382ac | FROGGS++ | src/6model/reprs/C (3 files):
start with align=0 as it will grow with attrs

When we compute the memory a CStruct, CPPStruct or CUnion will take, we will pick the biggest alignment of the struct members. Therefor it makes sense to not start with a value we might not reach. The previous code made a struct align to not less than 8 bytes which is problematic for inlined structs. This fixes RT #126675.		 19:43
FROGGS ahh, feels good to have anything to commit after that long time
19:43 vendethiel joined
hoelzro FROGGS++ 19:47
FROGGS :o) 19:49
jnthn \o/ 19:51
timotimo oh! 19:53
good work, FROGGS 19:54
i really don't feel like golfing libdigest into something that's somewhat minimal and still triggers The Bug
20:07 vendethiel- joined
jnthn finally has something to push that is along the lines of the thing he wanted to get working today :) 20:24
hoelzro I'm still unable to get a working MoarVM on Windows; it throws an access violation when trying to run the stage0 nqp.moarvm. Does that ring a bell for anyone? something with UTF-16 vs UTF-8 encodings, maybe? 20:25
jnthn hoelzro: Odd, I build/run Moar on Windows every day without trouble.
hoelzro: What compiler, bitness, etc?
hoelzro jnthn: 64-bit Mingw on Windows 8.1
I tried VS 2013 and 2015, same results 20:26
jnthn OK, 64-bit MSVC on Windows 7 here :)
o.O
hoelzro I'm wondering what I must be doing wrong!
jnthn Me too...
What Perl?
hoelzro strawberry
jnthn Though I don't think we steal too much out of its config...
FROGGS hoelzro: I'd suggest a 32bit strawberry toolchain 20:28
a 5.20 or older
hoelzro 32-bit all the way? compielr and all?
FROGGS hoelzro: strawberry ships a gcc and gmake you can use
just use that
hoelzro oh, really?
that's handy 20:29
retupmoca builds using strawberry perl and vs2015 without issue
hoelzro I'll try that, thanks FROGGS
FROGGS yes, install strawberry, and start configuring moar
hoelzro retupmoca: it seems like I'm the only one who has problems with this =/
that makes me wonder if mingw's compiler interferes with strawberry's
FROGGS quite possible
jnthn If you're going to use MSVC, then ActiveState's Perl builds are likely a better way to go, since afaik they too use MSVC too 20:30
FROGGS jnthn: you can choose AFAICT 20:31
activestate's perl lies about how it is made...
hoelzro if I could find my product key, I would just reinstall windows at this point 20:32
FROGGS it will tell it was built using MSVC, and if you install the mingw toolchain, it will decide to tell it was build using this one
hoelzro I think I may have just borked it
FROGGS :S
hoelzro I could complain about how hard it is to get a working toolchain, but I'm sure I would be saying the same about Linux if I never used it =/ 20:33
FROGGS true
jnthn I dunno what you're running into
FROGGS problem is that is easy to do too much there, and have something that conflicts 20:34
jnthn It's odd.
FROGGS jnthn: install mingw *and* strawberry, and you are likely screwed
so either use strawberry+gcc XOR strawberry+MSVC XOR activeperl+MSVC 20:35
hoelzro thanks for weighing in everyone; I feel like I do this every month =(
all this for testing Linenoise on Windows! 20:36
FROGGS hoelzro: if you wanna go the activeperl+msvc route, do steps 0,3,4 and 5 of github.com/rakudo/star/blob/master...ws-msi.pod 20:37
hoelzro oh, that looks handy! thanks, FROGGS! 20:38
I think MSVC would be ideal, because it's about as different as possible as my Linux dev env
FROGGS activeperl+msvc will probably also not interfere with mingw, just make sure the strawberry folder is renamed to something else, so perl.exe is from activeperl
aye
20:47 Peter_R joined
[Coke] hub.docker.com - rakudo-star has 48.8K pulls. 20:50
ww
hoelzro \o/
20:53 cygx joined
cygx in my more silly moments, I've been musing about creating a single-file windows build environment that budles busybox, tinycc, gnu make and a git clone based on libgit2 20:59
you need a unix-like build environment on windows? oh, it's just this file over here...
konobi there's free (as in beer) versions of visual studio these days 21:13
cygx that's not particularly unix-like, though ;) 21:16
21:35 colomon joined 22:35 colomon joined 23:33 tokuhiro_ 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!