Parrot 2.6.0 | parrot.org Log: irclog.perlgeek.de/parrot/today | Nopaste: nopaste.snit.ch:8001 | fix 'make html' (talk to Coke), merge gc_* branches, fix/replace/optimize hashing
Set by moderator on 3 August 2010.
kid51 In the other TT, I want to do refactoring on headerizer.pl. But that's second in the queue. 00:00
sorear: Melvin Smith was a Parrot contributor back in ancient times, like 2002. Somehow his personal copyright notice got stuck on many files in IMCC -- which poses licensing problems. Cf. trac.parrot.org/parrot/ticket/304 00:02
Austin Whiteknight: Merged and pushed. It doesn't work now. I don't know if the merge caused it, or if it's a timing problem.
whiteknight either way, anything is good
Austin gitorious kakapo master
00:04 wagle joined
Austin WTF? How did all that get commented out? 00:08
whiteknight ? 00:09
Austin kakapo_top.pir_tmpl is suddenly all commented.
That's why some stuff isn't working
whiteknight ...I may have done tht 00:10
that
Paul_the_Greek Question about .loadlib: None of the tests that do a .loadlib can find the specified file. 00:11
Austin What file?
Paul_the_Greek Is there something I missed doing?
sys_ops, bit_ops, debug_ops, etc. 00:12
Austin What os?
Paul_the_Greek Windows XP.
dalek kapo: 70e7a3e | austin++ | (27 files):
Paying the taxes on invisible methods and a bunch of other idiotic changes in 2.3-2.6.
kapo: 56c7fac | austin++ | :
review: gitorious.org/kakapo/kakapo/commit/...3014d3c5e0
Austin Look in $PARROT/config_lib.pir, check the value of the [build_dir] setting.
Paul_the_Greek It points to my top-level build directory. 00:13
Austin Well, I'm out of ideas. 00:14
Are you running the tests from that dir?
Paul_the_Greek Tests without a .loadlib work fine.
Yes.
The weird thing is that the .loadlib specifies, say, sys_ops, but the file is called sys.ops. 00:15
Austin Yea
But the dll is sys_ops.dll
Paul_the_Greek I have no such file.
cotto_work .ops files are postprocessed into C 00:16
Paul_the_Greek I have no ops DLLs at all.
00:16 aloha joined
Paul_the_Greek I have no bit.c, either. Hmm. 00:17
Perhaps I should remake my system.
Austin look in $PARROT/runtime/parrot/dyn* 00:20
Paul_the_Greek Remaking produced the files. Now let me try the test. 00:21
Much better.
And my new test works so much better. Thanks, Austin. 00:22
Austin Whiteknight: Another push, this one makes the bootstrap test run 00:26
tcurtis pmichaud: ping 00:30
dalek rrot: r48373 | mikehh++ | branches/tt1726_pmc_pod/t/codingstd/pmc_docs.t:
fix codetest failure - trailing spaces
 00:34
Chandon Good job Parrot. Somehow you just managed to deadlock signaling a condition variable. 00:40
Paul_the_Greek It appears that I need GMP to get bigints. 00:48
mikehh Paul_the_Greek: yes 00:49
Paul_the_Greek: what platform etc are you working on?
Paul_the_Greek Windows XP
Does this look good? cs.nyu.edu/exact/core/gmp/ 00:53
mikehh Paul_the_Greek: afraid I can't help there, I have been mostly working on Ubuntu distros for the last couple of years 00:54
Paul_the_Greek I'll check around.
mikehh Paul_the_Greek: that looks ok though 00:55
Paul_the_Greek I'll deal with it tomorrow. Take care, folks. 00:56
Austin Whiteknight: Beware of occurrences of Array::new in the code. These can be replaced by either [ ... ] for simple cases (no splat), or by calls to R*A.new( ... ) for cases with splats. Note that 'splat' in this case is spelled |@foo 01:11
or |<a b c>
01:14 dngor_ joined
cotto ~ 01:17
dalek rrot: r48374 | Chandon++ | branches/gsoc_threads (18 files):
[gsoc_threads] Just need a couple less explosions...
 01:24
01:25 plobsing joined
01:26 mj41_ joined, baest joined 01:27 workbench joined, Austin_Hastings joined, dmagnus_ joined 01:28 dalek joined, wagle joined
dalek kapo: d4fd7b0 | austin++ | :
review: gitorious.org/kakapo/kakapo/commit/...0a84371ee5		 01:31
01:35 dalek joined, dmagnus__ joined, Austin_Hastings joined 01:37 s1n joined, Austin joined 01:38 rurban_ joined, sjn joined, dmagnus_ joined, dalek joined 01:39 ascent joined, mikehh_ joined, ttbot joined 01:40 preflex joined 01:42 Austin_Hastings joined
kid51_at_dinner What are the HTML tags you use when you want to have some inline characters rendered in a monospace font? 01:43
GeJ kid51: <pre> ... </pre> 01:44
kid51 Will <code> work as well?
Austin_Hastings also <tt>
kid51 (this is inline, not whole paragraph)
01:45 slavorg joined
kid51 Well, at least on chromatic's blogs, none of those 3 work properly. Only <pre> changes the font ... but it imposes a paragraph. 01:49
mikehh opbots, names 02:00
dalek rrot: r48375 | jkeenan++ | failed to fetch changeset:
[codingstd] Merge tt1726_pmc_pod branch into trunk. Adds new file t/codingstd/pmc_docs.t to flag functions lacking documentation in .pmc files.
rrot: r48376 | jkeenan++ | branches/tt1726_pmc_pod:
Branch has been merged into trunk and is no longer needed at HEAD.
purl i already had it that way, dalek.
rrot: r48377 | jkeenan++ | tags/tt1726_pmc_pod-48342:
Branch corresponding to tag has been merged into trunk; tag may be deleted.
02:01 davidfetter joined
Austin kid51: www.w3schools.com/tags/tag_font_style.asp 02:02
GeJ kid51: an alternative would be to encapsulate your code inside a <span> et specify a style for it.
Austin But beware of style sheets that might override the formatting
02:12 mikehh joined
kid51 GeJ: Thanks, but that's too much work for a simple blog response: tinyurl.com/28qg9o5 02:17
Austin: Thanks for link. But perhaps the problem is that that blog's software only supports a limited number of HTML tags. 02:19
Austin Seems like it 02:21
purl somebody said Seems like it was creating a conflict somewhere, which then creates a requires condition
kid51 must sleep 02:24
purl $kid51->sleep(8 * 3600);
cotto must sleep 02:31
purl Sleep is for the weak.
cotto much better
mikehh All tests PASS (pre/post-config, make corevm/make coretest, test, fulltest) at r48377 - Ubuntu 10.04 i386 (g++) 03:18
t/op/exit.t - TODO passed: 6 in testf
darbelo mikehh: Do you, by any chance, test on non-x86 (and non-amd64) arches? 03:25
03:32 riddochc joined, janus joined
darbelo purl, msg bacek From some eyeball-and-wallclock testing it looks like the unshared buffers branch make the rakudo build go slower and eat more memory. My guess is heavy 'substr' use in PCT, but I'm not sure how to properly account for that. 03:33
purl Message for bacek stored.
cotto wasn't that the expected result? 03:34
sorear darbelo: have you optimized compact_pool yet? 03:36
riddochc I've been reading up on how to implement a new language targeting parrot. I just wanted to drop by and thank the people who have put so much work into the documentation. It's good stuff. 03:37
sorear Haha.
Great timing there.
cotto riddochc, thanks for the feedback. You can blame tcurtis for the shape of the squaak tutorial. He did most of the work to get it up-to-date for the 2.6 release. 03:38
sorear About two weeks ago.
Actually the documentation has been improved a lot in the last couple months 03:39
riddochc Nice. Maybe I'm crazy, but I'm really liking Clojure, except for the Java parts. ;) 03:41
tcurtis riddochc: If there's anything wrong with or that could be improved in the Squaak tutorial, please let me know. 03:42
riddochc I'm seriously thinking about Clojure for parrot. Maybe I'd call it 'pleajure'. I'll let you know, tcurtis. That's specifically what I've been looking at.
Actually, I doubt I'll have the time or energy to do it. But it's a fun idea, anyway, and an excuse to see what's new with parrot. 03:45
tcurtis riddochc: I feel obligated to warn you that Parrot's threading/concurrency support is rather lacklustre. Although one of the GSoC students has been working on it, so maybe that will change soon. 03:46
riddochc tcurtis: Good to know. Oh, well. Perhaps it's a good idea in the long-term. 03:48
03:50 GeJ joined
mikehh darbelo: I only have x-86 hardware at present, and can only test on that 03:51
darbelo: I also had problems installing a a VM so need to reboot to test on another distro etc, so I stick to various versions of Ubuntu (usually the latest) 03:54
darbelo mikehh: Not a problem, I asked out of curiosity, mostly. I know we have testers on various OSes, but I don't really what variety we have on the hardware front. 03:58
sorear: Not yet, really. The only 'optimization' I've done to the gc is remove the flag checking and twiddling. 04:00
The compact_pool logic is still as... sub-optimal as it is on trunk. 04:01
04:02 TiMBuS joined
mikehh darbelo: darbelo: TapTinder does some nice testing - tt.taptinder.org/ 04:03
bah getting too late for me, my typing is getting erratic 04:04
darbelo sorear: But I'm not sure that acounts for most of the runtime penalty, really. Or the memory, for that matter.
mikehh: Yeah, but that's still mostly x86 and x86-64 AFAICT. 04:06
sorear darbelo: AFAIU, the only real reason to bother with unsharing strings is compact_pool improvements 04:07
mikehh darbelo: I know kid51++ has an older Mac (pre x86) that he test on and AndyA++ does some solaris testing
riddochc It's getting pretty hard to find hardware that isn't x86, x86-64, and ARM these days.
darbelo It depends on your environment. In some places, SPARC boxes are still pretty common. 04:09
sorear riddochc: it's getting pretty easy to get *very* good emulators these days 04:11
darbelo And there's those chinese mips64 laptops I can't remember the name of right now. 04:12
Austin Loongson 04:13
mikehh I have some $work scheduled later in the year on a new IBM Power 780, maybe I can con^W, suggest that we can test parrot on that 04:16
darbelo Austin++ # Yep, those. 04:17
riddochc sorear: that's true, the emulators are decent.
mikehh has anyone done any parrot testing on ARM? 04:19
riddochc certainly good enough for testing Parrot.
darbelo mikehh: To my knowledge, we don't have anyone *building* on ARM on a regular basis.
It worked at some point, and it probably still does. 04:21
But we don't have regular testers, so we don't know for sure.
riddochc Aren't a lot of the phones out there these days running ARM? 04:22
sorear Yes, but most phones don't have access to the several trillion Hz-seconds required to build Parrot 04:23
darbelo And, parrot doesn't do cross-builds at the moment. 04:24
riddochc That's unfortunate. Though understandable - cross compiling is tricky stuff. 04:25
darbelo Enabling cross-development was a big part of the parrot-on-RTEMS GSoC, but I don't know how far along that is.
sorear Especially when the build system involves running tools written in PIR
darbelo That can be solved by judicious use of a host-side parrot. 04:26
plobsing we would have to distinguish between $(HOST_PARROT) and $(TARGET_PARROT) in the makefile. we don't do that ATM. 04:27
riddochc Well, I know some people who specialize in cross-builds, but they don't really have a lot of copious free time to help out with parrot, sadly. 04:29
mikehh We still have problems relating to a built parrot and installed parrot
darbelo And since we're asking for ponies, our build could be made smarter about some of the bootsrap steps too. 04:30
riddochc Ouch.
mikehh darbelo: very much so
sorear riddochc: Fortunately, we have someone being payed to do this thankless job. 04:31
dalek kapo: da37fc8 | austin++ | (5 files):
Fixed accidental delete of Parrot:: namespace sub exports
mikehh anyway gotta go, it's 5:30 am here and I seriously need a nap 04:32
darbelo sorear: Well, we have some student's getting paid to do *some* thankless jobs. I don't know of anyone outside of the GSoC getting paid to hack on parrot. 04:33
riddochc fwiw, I think the parrot and perl 6 projects have managed to avoid the biggest problem in the industry - being in too much of a hurry to do anything well. I know you've gotten a lot of flack for the timelines of the projects, but I think you're better off doing it this way. 04:35
I've had enough of barely-working software getting shipped too early. So, cheers to parrot and perl 6. 04:36
Have fun, everyone. I'm off for now. 04:40
04:40 riddochc left
dalek rrot: r48378 | plobsing++ | branches/dynop_mapping/src/pmc (2 files):
rework OpLib and Opcode to fit new model better.
 04:54
rrot: r48379 | plobsing++ | branches/dynop_mapping/t/pmc/pmc.t:
fix test - OpLib now requires an initializer
Austin msg whiteknight Just pushed some fixes that get the UnitTest library passing bootstrap. Replaced all method with our-method, which is probably stupid in the long run. 05:25
purl Message for whiteknight stored.
dalek kapo: 0075f45 | austin++ | (108 files):
Got Testcase's working (bootstrap). Replaced all methods with our-method
 05:27
Austin msg whiteknight prove -r --ext=nqp -e parrot-nqp t/bootstrap # this works for me - all bootstraps pass. 05:45
purl Message for whiteknight stored.
dalek kapo: a46ea16 | austin++ | (3 files):
Got all bootstrap tests working (String::isa)
 05:51
Andy Thanks moritz++ perlbuzz.com/2010/08/im-glad-to-hea...-slow.html 05:54
dalek rrot: r48380 | plobsing++ | branches/dynop_mapping (3 files):
check for core_ops (doesn't occur as a loaded dynoplib).
 06:05
06:21 uniejo joined
dalek TT #1735 created by plobsing++: Pmc2c line numbering confuses gdb 06:21
TT #1735: trac.parrot.org/parrot/ticket/1735
06:30 jsut_ joined 06:36 LoganLK joined
dalek rrot: r48381 | plobsing++ | branches/dynop_mapping/src/pmc/oplib.pmc:
include appropriate header for fetching core_ops lib
 07:14
rrot: r48382 | plobsing++ | branches/dynop_mapping/t/pmc/oplib.t:
update oplib tests to use initializer
 07:30
07:54 Casan joined 08:45 arnsholt joined 08:46 arnsholt_ joined 08:50 AndyA joined 09:27 bacek joined 09:35 aloha joined 09:37 rurban_ joined 09:41 cognominal joined 10:16 clinton joined 10:23 AndyA joined
dalek kapo: 9bac5a4 | austin++ | (2 files):
Checkpoint. Trying to get FileSystem up. DepQ may not be running
 10:42
kapo: 5f2617b | austin++ | src/ (24 files):
Got UnitTest library restored, at least somewhat. Reverted damage to Matchers
10:55 GeJ_ joined 11:07 GeJ joined, luben_work joined 11:11 lucian joined 11:14 slavorgn joined 11:45 wagle joined 11:46 rurban_work joined 11:50 whiteknight joined 11:59 clinton joined 12:09 rurban_work left 12:15 kid51 joined 12:41 lucian_ joined 12:44 lucian__ joined
Coke riddochc++\\ 13:01
riddochc++ #even 13:02
13:04 Andy joined
Coke is TT#1735's bug that the line number needs to be fixed? 13:09
(because there's already a ticket for that.
... that I cannot find. ah well. 13:10
Andy Oh sure, Coke. Make ME do the modeline development. 13:16
Do we have an existing .t that would be good to check for it in?
13:18 bluescreen joined
Coke codingstd/*coda*.t 13:24
which has a check for existing ones.
but I'd make it a patch, because it's going to cause "make codetest" to fail.
13:41 bluescreen joined 13:48 plobsing joined
bluescreen ~prod support 13:52
jdv79 i hadn't realized how much nqp had progressed since last i looked. nice work. 14:03
14:04 bubaflub joined 14:07 ruoso joined 14:25 Paul_the_Greek joined
Paul_the_Greek Hello everyone. 14:25
Does anyone have an idea which file has to be where in order for configure to decide I have GMP installed?
moritz typically the GMP development files 14:27
ie headers
Paul_the_Greek So gmp.h has to be somewhere? Do you know where? 14:28
moritz where the compiler can find it.
Paul_the_Greek And a .dll somewhere, I presume (I'm on Windows). 14:29
moritz it's /usr/include/gmp.h for me
Paul_the_Greek I would think putting it somewhere in the MinGW tree would do the trick.
moritz Paul_the_Greek: write the simplest program you can think of that uses GMP. Try to compile it. If it works, your installation is fine 14:30
or maybe use an example program from the gmp documentation
Paul_the_Greek Configure doesn't see it. I'll keep futzing.
moritz if you want to burn your time, sure go ahead fuztzin 14:31
Paul_the_Greek Trying to figure out how to compile a program is exactly as difficult as trying to figure out where Configure wants it. 14:32
That's because I have absolutely no idea where to put the GMP files.
moritz IME that's not the case. If you try to compile stuff, the compiler actually says what's missing 14:33
whereas configure.pl just says "no."
though maybe with verbose output...
Paul_the_Greek It's installed with Strawberry Perl, too, but that doesn't help. 14:35
I'll try what you suggest, but I don't know why I couldn't compile a program if the .h file is in include and the .dll file is in bin. 14:40
Coke the test configure runs is to add -lgmp to the compile options and try to run a c program. 14:41
moritz that's what the compiler is going to tell you :-)
Coke there's no check for any libs or executables. just "does this program compile and generate the output I expect." 14:42
Paul_the_Greek: are you adding the lib to the command line for the compiler?
Paul_the_Greek Sorry, I haven't tried to compile a standalone program using GMP. I will do that. I just thought someone would know where to install GMP on Windows. 14:44
Coke I'd ask jnthn or particle. 14:50
but If this was unix, I'd just install it, and it's either in the right spot, or I'd have to tell parrot where it was when I configured. 14:51
14:54 mikehh joined
Paul_the_Greek I can't find a Windows installer, only a gzipped file. 14:55
Okay, compiling with -l has no problem with the include, but then complains about entry points. That makes sense. 14:59
When I use the -lgmp option, what file is it going to look for? 15:03
15:05 theory joined
Andy Coke: Or I can both update codingstd/*coda*.t AND update the codas. 15:05
Coke Andy: that would be ideal, yes. 15:06
I was trying to provide smaller work-chunks.
Andy++
Andy I DO NOT NEED SMALL WORK CHUNKS
dafrito haha <3 15:10
mikehh Hey Andy, you see we have this small problem with parrot, it is called Garbage Collection, gc foir short, and if you are looking for a smallish chunk of work .... 15:11
Andy mikehh: I NEED WORK CHUNKS THAT I AM SUFFICIENTLY SMART ENOUGH TO HANDLE.
mikehh Andy: ah well it was worth a try :-} 15:12
Paul_the_Greek Can you tell me what the -lgmp option does and what file it looks for? There is no documentation on -l from gcc --help --verbose. 15:15
Coke -l<library> 15:17
atrodo it tells gcc to link to the gmp library, which under unix would be libgmp.a (pretty sure)
Coke gcc.gnu.org/onlinedocs/gcc/Link-Opt...nk-Options
Paul_the_Greek When I installed GMP on Windows, I got libgmp-3.dll and libgmp.la 15:18
Oh, that documentation helps, perhaps. Thanks Coke. No such option display with gcc --help --verbose. 15:20
atrodo the key is that those two files should be found in ld's library path. You can try a "export LDFLAGS=<path>" before you configure 15:22
Coke atrodo: he's on windows. 15:23
atrodo hehe, i saw mingw and went the extra mile to assume bash. I would assume mingw checks env variables as well 15:24
Paul_the_Greek Success! 15:29
purl well, success is finding king size papers
Paul_the_Greek That was much harder than it needed to be. 15:30
Thanks, folks.
mikehh All tests PASS (pre/post-config, make corevm/make coretest, test, fulltest) at r48382 - Ubuntu 10.04 amd64 (gcc with --optimize) 16:19
16:22 theory joined 16:27 robin-gvx joined
cotto_work ~~ 16:55
atrodo hey cotto_work
purl cotto_work is, like, always at work or cotto's employed alter ego
atrodo it makes it so difficult to actually say hello to someone sometimes
cotto_work you mean purl? 16:56
atrodo Yep 16:57
Paul_the_Greek Well, that was lovely, but it's GMP 4.1, which is worthless. 17:10
No chance anyone here runs on Windows, is there?
17:22 bubaflub_ joined 17:37 rurban_ joined
Coke Hello, Reini. 17:56
17:57 theory_ joined
dalek kudo: 22a11d0 | moritz++ | (22 files):
Merge branch 'master' into match-call

   src/core/Match.pm		 18:00
kudo: 77d4761 | moritz++ | src/core/Match.pm:
fix wrongly resolved merge conflict
kudo: 9412433 | moritz++ | src/ (3 files):
!~~ now topicalizes too
kudo: 87c82c2 | moritz++ | src/ (4 files):
Merge branch 'match-call'

sets the topic for negated smartmatch (!~~)
18:08 bubaflub joined 18:13 Paul_the_Greek joined 18:30 robin-gvx joined
cotto_work #ps in 118 18:33
dalek kudo: fcf4f36 | moritz++ | (3 files):
adverbs for m//
 18:35
18:39 luben_work left, Chandon joined 19:27 smash joined
smash hello everyone 19:27
moritz smash! 19:28
19:34 lucian joined 19:36 theory joined 19:38 mikehh joined 19:40 tcurtis joined
Austin pmichaud++ #'multi' works in nqpr 19:58
x
dalek kapo: b9c38b8 | austin++ | (4 files):
Got FileSystem working with multi methods (pmichaud++)
 20:00
cotto_work It's nice to have multis in nqp 20:01
dalek rrot: r48383 | Chandon++ | failed to fetch changeset:
[gsoc_threads] Merge from trunk.
 20:07
smash nominations for next parrot foundation board of directors are up, as soon my e-mail hits parrot-members 20:20
20:21 chromatic joined 20:25 theory joined 20:26 AndyA joined
mikehh #ps time 20:30
20:32 AndyA_ joined
GeJ Bonjour everyone. 20:40
chromatic aloha
Coke ho. 20:41
tcurtis Hello.
cotto_work mro? 20:43
purl mro is the Digital sitecode for Marlboro, MA or Method Resolution Order or www.python.org/2.3/mro.html or one solution to the 'Diamond Problem' at en.wikipedia.org/wiki/Diamond_problem or Method Resolution Order or a pragma, see the 5.10 docs for details
cotto_work Paul_the_Greek: ^
chromatic When you dispatch a method, which class's variant of that method gets called
Coke MMDRO! 20:44
moderator Parrot 2.6.0 | parrot.org Log: irclog.perlgeek.de/parrot/today | Nopaste: nopaste.snit.ch:8001 | merge html_cleanup (talk to Coke), merge gc_* branches, fix/replace/optimize hashing 20:45
darbelo chromatic: In my case: (0) Point at the area where we need to dig. (1) Explain what we need to do (2) Give a 'easy' way to determine success of the task. (3) Prod periodically. 20:46
(iv) Make me stop talking on the wrong chanel. 20:47
Coke msg bacek I'll buy you a beer if you finish gc_massacre! 20:56
purl Message for bacek stored.
21:03 whiteknight joined
cotto_work Shipping might be pricey. 21:12
chromatic seen plobsing
purl plobsing was last seen on #parrot 16 hours, 45 minutes and 9 seconds ago, saying: we would have to distinguish between $(HOST_PARROT) and $(TARGET_PARROT) in the makefile. we don't do that ATM.
whiteknight hash_allocator branch is a priority? 21:13
dalek rrot: r48384 | darbelo++ | branches/unshared_buffers/src/string/encoding/utf8.c:
Make utf8 not depend on str_new.
rrot: r48385 | darbelo++ | branches/unshared_buffers/src/string/encoding/utf16.c:
Make utf16 not depend on str_new.
rrot: r48386 | darbelo++ | branches/unshared_buffers/src/string/encoding/ucs2.c:
Make ucs2 not depend on str_new.
rrot: r48387 | darbelo++ | branches/unshared_buffers/src/string/encoding/ucs4.c:
Make ucs4 not depend on str_new.
rrot: r48388 | darbelo++ | branches/unshared_buffers/src/string/encoding (5 files):
Add smarter handling of constant and external strings.
smash & brb 21:17
21:22 luben joined
Util Paul_the_Greek: where did you find 4.1.3 binaries? 21:22
luben hello guys 21:27
Paul_the_Greek Util: cs.nyu.edu/exact/core/gmp/ 21:28
cotto_work hi luben 21:30
luben I have seen that hash table rework is a high priority task. I was looking at he hash.c code and have a pretty good understanding of ist internal logic
I have some quick optimizations that could get us around 3-4% 21:31
cotto_work luben: by all means, nopaste a patch or file a ticket 21:32
luben but the current code is ok as it is
I have some design questions 21:33
dukeleto luben: that sounds awesome, maybe you can create a svn/git branch for that work
luben if we want to gradually redesign the hash logic, this optimisations will be reverted (that depends on the new design)) 21:34
whiteknight luben: what kind of redesign are you thinking? 21:35
luben Also I would like to talk about the overall principles
Paul_the_Greek I'm interested in listening in on the hash issues. 21:36
luben but I am not quite sure that it will not break parrot in other parts
dukeleto luben: that is what our test suite is for :)
luben that poke with the hash internals
Util Paul_the_Greek: Any reason not to jump ahead to GMP 5.0.1? 21:37
dukeleto is surprised to hear that GMP 5.x is out
Paul_the_Greek I don't think so. The BigInt tests check for >= 4.1.4
luben so, first, one quick question: I have seen that hash_table struct has a pointer to its container
and on each update it checks if the container is defined as constant and refuse the operation in such a case 21:38
chromatic Seems backwards, doesn't it? 21:39
luben It is used in such a manner by OrderedHash.pmc and Namespace.pmc
My logic is that enforcing the policy of the container is work of the container, not the underleing data structure 21:40
chromatic Agreed.
dukeleto bubaflub: things proceeding? stuck on anything? 21:41
bubaflub dukeleto: yeah, i'm actually hacking on it as we speak
dukeleto bubaflub: goodly :)
bubaflub dukeleto: just a bummer knowing that i won't make my original target
NotFound luben: killing that pointer to the container will be good.
dukeleto bubaflub: it happens. don't think about that, and let's see if we can get out-of-dir compiling working, that is a big step in the right direction 21:42
luben the fast optimisations: we could remake hash_mark_keys,hash_mark_values and hash_mark_both to walk lineary the buckets
bubaflub dukeleto: i still want to finish what i set out to do, but it just might take a while
luben and we destinguish occupied buckets by key!=NULL 21:43
chromatic luben, I think that only works if the bucketlist/freelist are linear.
luben yes. but most of the time they are (if we do not have a lot of deletes)
21:44 bacek joined
dukeleto bubaflub: that is what everyone wants. Let's try to get something tangible by the hard-pencil date, and then we can work on this a bit more leisurely :) 21:44
luben hash buckets are allokated lineary from the top to the tail
cotto_work bubaflub: whiteknight didn't get his original gsoc target either. Don't worry too much about it.
bubaflub dukeleto: indeed. the first half is tangible - i've got the secret sauce to do a minimum libparrot for cross-compiling.
21:45 aloha joined
dukeleto bubaflub: you will pass as long as everyone thinks you tried your best. We all know you are dealing with some hairy internals across two very different projects, which ain't easy. 21:45
luben As I said that I am not so sure about long term viability of this aproach
The other idea i have is a little more complex 21:46
mikehh Paul_the_Greek: gmplib.org/
luben it is not local optimisation, but a whole (gradual) rework of the hash data structure
bubaflub dukeleto: well, the good news is i'm unstuck from that problem i reported on tuesday. 21:47
Paul_the_Greek mikehh: They have no binaries for Windows. Not sure they have binaries at all.
bubaflub dukeleto: Configure.pl now will handle making the dir's necessary in the build directory. Now, gotta hack the Makefile to reference the source dir where appropriate
chromatic Reworking the whole hash seems reasonable.
mikehh Paul_the_Greek: just the source
luben My curiousity was provoked by 75% rule that was removed by chromatic
smash back 21:48
luben because there is such a rule in hash tables design
Paul_the_Greek Yes, but it can't be built on Windows. At least not that I can figure.
luben but this is not our case (we use separate chaining with pre-allocated space) 21:49
chromatic The rule as I understood it was "Reallocate when you're 80% full"
21:49 bubaflub joined
mikehh Paul_the_Greek: maybe you should install Ubuntu 10.04 LTS, it can be installed within Windows if you like 21:49
luben yes, not it does not concern separate chaining hash implementations
s/yes, not/yes/ 21:50
Paul_the_Greek mikehh: Ouch. I'm a Windows guy. Anyway, we have to get binaries for our users, no?
luben And I nave a rather radical idea
in the situation now we have the same size of hash buckets and hash_indexes 21:51
mikehh Paul_the_Greek: I know saome of the rakudo developers use Windows, jnthn for example, you could check with him maybe
Paul_the_Greek mikehh: I will check. 21:52
tcurtis needs to get around to figuring out how to get Parrot's configure to recognize his libgmp installation at some point.
luben indexes store a pointer to a bucket: i = hash(key)%N = index[i] = *bucket 21:53
could we eleminate the entire hash indexes and address directly the buckets
chromatic Makes sense.
dukeleto tcurtis: you may need to add some directories to LD_LIBRARY_PATH if you installed libgmp in a non-standard place 21:54
luben we will have a 2/3 of our size
and my profiling the previous 2-3 days have shown that the size of the data structures is primal factor to hash table performance 21:55
example: just adding additional field in bucket struc could lower the performance with 5-6%
chromatic Rakudo creates some 44,000 hashes to start up. 21:56
Paul_the_Greek Wow.
luben yes, I have seen in the callgring/cachegrind profiles... it's huge
my concerns are: 21:57
there are some other pieces of parrot that seem to poke with current hash_indexes:
OrderedHash and args.c 21:58
(probably for performance reasons)
chromatic Right.
luben if we eleminate indexes they could use straight addressing and have the same benefit 21:59
chromatic I'm all for this. 22:00
Paul_the_Greek So there is a block of buckets and then also a vector of indexes into them?
luben if we make such a elemination, we have something that is known as "coalesced hashing" or "coalesced chaining"
here is quick overview en.wikipedia.org/wiki/Coalesced_hashing
Paul_the_Greek I believe Lua uses coalesced hashing. 22:01
luben it is very efficent memory-wise, with very fast inserts and lookups
but with slower deletes 22:02
and more complex expanding algorythm
Paul_the_Greek I implemented a coalesced hash table for a language I was working on. You can take a look at it if it would be of any help.
luben I have some prototypes 22:03
just to get the idea
Paul_the_Greek It also moves an entry that is not in its home position if a new entry comes along with that home position.
luben but my understanding is thet we should gradually morph our hash in the desired type 22:04
chromatic If we can get rid of container before the 2.7 release, we should see modest performance and memory improvements.
Paul_the_Greek luben: How can I send you a small chunk of text, the description of my hash table? 22:05
cotto_work Paul_the_Greek: nopaste
Paul_the_Greek nopaste?
purl nopaste is nopaste.snit.ch (works with the script in $_PARROT/tools/dev/nopaste.pl) or paste.scsys.co.uk or www.extpaste.com or gist.github.com or App::Nopaste or codepeek.com/paste/ or (: pastebot)
luben actually it is not a lot of work: just mouve i check and exception raise from hash.c to corresponding PMCs 22:06
chromatic Sounds easy to do.
luben yes, it is used only in src/pmc/orderedhash.pmc and src/dynpmc/dynlexpad.pmc 22:07
also, if we move now to liner scan in mark functions, they will work as is without indexes 22:09
chromatic Can you make two patches for those?
luben a could send a patch tommorow
yes
I have done it and revered, just to explore other routes 22:10
chromatic Thanks! 22:13
bubaflub dukeleto: i've confirmed that my changes work; all that remains for out of directory building is fixing the Makefile to point to the source dir explicitly. i'll be working on it a bit tonight.
luben after elimination of bucket indexes (this will gain us some performance) we could gain more performace if we need it (compromising memory) 22:14
with allocation of overflow area
elemination bucket indexes will speed the hash put/get functions because we will have one less indirection 22:16
and will reduce the memory size
I will send the patch for mark functions in an hour 22:17
Also 1 question: full hash revamp is not planned for 2.7 release 22:20
chromatic Right.
dalek rrot: r48389 | darbelo++ | failed to fetch changeset:
Sync with trunk.
luben ok:) because I am plannong a 2 week vacation from next monday 22:21
tcurtis What's the proper way to make a PAST::Val produce a Float rather than an Integer? :returns<Float> doesn't appear to work. 22:23
22:40 bacek joined, aloha joined
luben ok, here is a patch for linear scanning of the buckets on gc_mark : pastebin.com/m6KAsbw5 22:44
it gives me a moderate improvement in rakudo startup benchmark: from 690ms to 660ms 22:46
by my calculation it improvement of 4% 22:48
chromatic Let me use Callgrind. 22:49
luben my benchmark is : $time parrot --hash-seed=D3C21BCF466DA1 ../rakudo/perl6.pbc -e '' 22:50
dukeleto luben: are you doing multiple runs or is that % improvement from a single run? 22:52
luben we should consider also CPU cache misses, not only instructions executed
yes, it is not from a single run
chromatic I'll run Cachegrind too.
luben first run is discarded because of the disk IO 22:53
then I make 10 runs
and discard the extremes 22:54
dukeleto luben: sounds great, just checking :)
dalek rrot: r48390 | chromatic++ | trunk/src/call/args.c:
[PCC] Tidied code; no functional changes.
rrot: r48391 | chromatic++ | trunk/src/pmc/hash.pmc:
[PMC] Reordered some Hash code for clarity.
luben and make an average of the rest
chromatic, I use this: valgrind --tool=callgrind --cache-sim=yes --branch-sim=yes
valgrind Release 3.5.0 22:55
chromatic I haven't used --cache-sim, good idea.
luben and than analyze data in kcachegrind
this simulates an avarage CPU cache 22:56
you could set as an options the exact parameters of the CPU cache hierarhy, but I hope that they have picked a reasonable defaults 22:57
22:57 jsut_ joined
Paul_the_Greek How would I contact a Rakudo person named jnthn? 22:59
luben jnthn?
purl i think jnthn is giving a lecture at the uni friday afternoon i think
cotto_work ping him here or ping jonathan in #perl6 in freenode
tcurtis cotto_work, Paul_the_Greek: I'm fairly certain jnthn is jnthn both here and in #perl6. 23:00
Paul_the_Greek join #perl6
Oops.
cotto_work for some reason I thought otherwise
my mistake
chromatic So far that patch looks like an improvement, luben. 23:01
Paul_the_Greek When you guys leave a message for someone here, where does it show up? 23:02
chromatic msg Paul_the_Greek Here's how purl works. Respond to me in channel. 23:03
purl Message for paul_the_greek stored.
dalek parrot: e9bcd4a | (Jonathan "Duke" Leto)++ | html/docs.html:
Link to the PGCon website, which has the presentation and audio
chromatic luben, it's an improvement here of modest performance (0.657% on Callgrind, about that for Cachegrind). 23:06
23:06 aloha joined
luben it is possible, my calculation was based on wall clock (64bit linux) 23:13
23:13 lucian joined
dukeleto luben: yeah, wall clock isn't very useful for benchmarking purposes, since it is a function of the load of the machine 23:14
luben there is no other load 23:16
or not considerable load
Austin Paul_the_Greek: say something 23:19
purl something
Paul_the_Greek Hey ho!
Austin And purl says, "You have a message!"
Util Paul_the_Greek: Did you try to build GMP with MinGW+MSYS, or just plain MinGW ? 23:20
Austin Then you say, "purl, messages"
Paul_the_Greek Plain MinGW. Except I couldn't convince myself that the makefile would even work on Windows.
I must admit I didn't try too hard. 23:21
mikehh Paul_the_Greek: you can try a ping, as in name: ping, which may gender a response, or otherwise msg 23:23
Austin engender or generate
(or produce or yield or elicit) 23:24
mikehh I think I meant to say render :-}
Austin You didn't, really.
Yes! That worked. 23:25
23:26 hercynium joined
mikehh Austin: early hours of the mornin' for me at present :-} 23:26
Austin :-)
clock 23:27
clock?
purl Austin: LAX: Tue 4:27pm PDT / CHI: Tue 6:27pm CDT / NYC: Tue 7:27pm EDT / LON: Wed 12:27am BST / BER: Wed 1:27am CEST / IND: Wed 4:57am IST / TOK: Wed 8:27am JST / SYD: Wed 9:27am EST /
mikehh am on BST 23:28
Austin Yoiks
mikehh at least so they tell me 23:29
Austin We are EUROPEANS! You may take our currency, but you'll never take our ... TIME ZONES!!!
mikehh you colonials don't seem to appreciate these things ... 23:31
Austin Heh. Don't point the finger at us. Did you see the Sovie^W Cissi^W Russians are talking about legislating away about 2/3ds of theirs? 23:32
(Admittedly, we only use two...) 23:33
mikehh right, like when I was in Hawaii , that time... 23:35
luben chromatic, it is a step in the direction of eleminating hash_index and the code is more simple, but it assumes certain properties of the bucket allocation (linear from the top) 23:36
chromatic found a 4.383% performance improvement in Parrot startup.
Austin Yeah, they don't even use clocks, I think.
But all the ones they do have are set to Pacific time, so they know when the next planeload of tourists is taking off.. 23:37
Ahh, multi-methods. You were doing so well until I tried to actually CALL you... 23:38
luben I think cachegrind defaults are very conservative (something like P3). now I am running some passes to compare the default with the cache hierarchy of current CPUs 23:40
sorear chromatic: Do you feel TODO is a failing test? 23:47
chromatic I only TODO expected failures.
For example, an unimplemented feature I want to test anyway gets a TODO. 23:48
sorear Also, what do you recommend we do about the 2 hour Perl 6 testsuite, it's driving me up the wall
chromatic NQP-rx profiling and optimization. 23:49
Paul_the_Greek Take care, folks. 23:50
chromatic My intuition says that there are a handful of hotspots there that need optimization at the PIR/PCT level, not the Parrot level. 23:53
(Setting aside the fact that improving Parrot hashing, register allocation, and GC performance will improve Rakudo dramatically.)
23:54 ruoso joined
chromatic Rakudo's runtime isn't super speedy, but something in the parsing stage has the feel of an algorithmic problem. 23:55
ttbot Parrot trunk/ r48393 i386-linux-thread-multi make error tt.taptinder.org/file/cmdout/365428.txt ( tt.taptinder.org//buildstatus/pr-Pa.../rp-trunk/ )
23:59 Psyche^ 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!