freenode/clasp - IRC Chatlog

15:07:10 drmeister Hi

15:08:34 drmeister I added this debugging code that monitors correctness of closures.

15:09:41 drmeister It works fine and no problems are identified unless i salt them in.

15:10:17 drmeister So I turn off everything and then start slime - it segfaults when I leave sldb - argh.

15:10:38 drmeister So everything is being turned back on to find the new problem.

15:31:54 Bike tracing jit-finalize-repl-function and compiling inline has some confusing output. it seems to finalize functions named initialize-instance and some functions related to cleavir ir thousands of times.

15:32:34 Bike when is jit-finalize-repl-function supposed to be called, exactly? whenever a function is compiled for the image?

17:32:48 drmeister Yes, it is called whenever a function is compiled for the image.

17:33:09 drmeister My problem with slime was not a problem. I was running an old version of bclasp (sigh).

17:33:16 drmeister All is well.

17:33:32 drmeister I am ready to move on.

17:34:11 drmeister Turning off debug code and rebuilding (sigh)

17:35:18 drmeister My fingers bang out "~/Development/clasp/build/boehm/iclasp-boehm" faster than my brain can process.

17:40:12 Bike still hangs. oh dear.

20:20:41 Bike i'm running with echo repl read and it seems to be proceeding, just very slowly. which doesn't make any more sense, really.

20:39:21 drmeister Hello

20:39:49 drmeister Bike: Do you have any more insight?

20:40:02 drmeister shiho: I read the smirnoff force field stuff - it's really easy.

20:40:16 drmeister We could implement it in a couple of days.

20:40:17 shiho good news

20:40:17 Colleen shiho: drmeister said at 2018.01.08 22:52:57: I pushed the changes

20:40:33 drmeister https://github.com/openforcefield/smirnoff/blob/master/The-SMIRNOFF-force-field-format.md

20:40:51 drmeister They are basically eliminating atom types (which I like) using things like this...

20:41:03 drmeister <Bond smirks="[#6X3:1]-[#6X3:2]" k="820.0" length="1.45"/>

20:41:09 drmeister XML - bleh

20:41:28 drmeister Better: (bond :smirks "[#6X3:1]-[#6X3:2]" :k 820.0 :length 1.45)

20:42:16 drmeister The smirks code has :1 and :2 which tag the atoms that are recognized by the smarts code and then the force field parameter is assigned if there is a match.

20:42:54 shiho I see. :)

20:43:12 drmeister I will implement an esrap SMIRKS parser and then with that we can implement this with a couple of lines of code.

20:43:40 drmeister These are basically programs that operate on atoms/bonds/angles/dihedrals to assign parameters.

20:44:02 Bike hum. it just finished. only took like an hour to go through one file.

20:44:10 drmeister "Children playing with children's toys"

20:44:25 drmeister WTH?

20:44:35 Bike obviously that means my optimization is not, but i don't see how that can be

20:44:54 drmeister But it's producing the correct result?

20:45:04 Bike well, it finished compiling inline.lisp.

20:45:06 drmeister Is the interpreter being used/abused?

20:46:07 Bike not that i'm aware of.

20:46:46 drmeister Try (time (compile-file "sys:kernel;cleavir;inline.lisp")) Does it take the same amount of time? That should produce a count of how many times the interpreter is invoked.

20:46:53 drmeister If its' not the interpreter then profile it.

20:47:00 drmeister Something crazy is going on.

20:51:47 drmeister Of course naive profiling of the compiler is a mess - but you can run the 'count-calls' filter on it and look for suspicious things.

20:53:47 Bike well, since it's inline it's actually loaded rather than compile filed

20:55:35 Bike but i'l time it

20:55:48 Bike this is from my method change, and that somehow causing interpreter use would be, uh, surreal

20:56:44 drmeister But the observation that it is working - but really, really, really slowly is very perplexing.

20:57:56 drmeister Also - try using ./cclasp-boehm -s

20:58:15 drmeister The -s will dump a list of debug flags and whether they are turned on or not.

20:58:27 drmeister Maybe there is some bad interaction with a debug setting.

21:00:12 Bike this is bclasp

21:00:14 drmeister Bike: Do you want a big heavy wooden desk?

21:00:29 Bike heh?

21:00:46 drmeister I just got a standing desk at home and I need to get rid of the old wooden one I have.

21:01:06 drmeister My first choice would be to give it to you.

21:01:22 drmeister Second choice is to find some non-murderer on Craigslist who will take it away.

21:01:35 drmeister Someone always gets murdered in a Craigslist interaction.

21:02:25 Bike i found my apartment on craigslist, so that's a bit worrying

21:02:39 Bike but no, while in the abstract i'd like a wooden desk, i wouldn't have anywhere to put it

21:02:45 drmeister Anyway - tell me soon offline if you want it - otherwise it's kill or be killed for me.

21:03:00 drmeister Ok.

21:03:45 drmeister bclasp honors all of the debug settings.

21:11:40 Bike well, it has use symbols in global array, debug recursive allocations, meter allocations, and debug release defined.

21:11:48 Bike running iclasp-boehm though.

21:12:10 drmeister Those are all innocuous - so it's not that.

21:12:52 drmeister Uh - how do you reproduce the behavior in bclasp?

21:15:02 Bike load cleavir, load auto compile, load inline. though given how diverse inline is it's possible that "the behavior" is just cleavir being much slower to compile.

21:15:37 drmeister You said an hour to load inline.

21:15:48 drmeister (time (load "sys:kernel;cleavir;inline.lsp"))

21:15:55 drmeister That would be an hour?

21:16:47 drmeister You might be on track to discover something that is slowing down cclasp compilation because you might be exacerbating it.

21:17:18 Bike yes. i'm running that form right now and waiting to see how long it takes.

21:17:59 drmeister I added code to get more detailed timing information on every form and every stage of cleavir compilation.

21:18:53 drmeister From my notes: "Set cmp::*debug-compile-file* to T to get timing on forms compiled by cclasp"

21:19:24 drmeister It should/could work here because it is cleavir that is doing the compiling

21:20:25 Bike does that work without compile-file?

21:20:46 drmeister Right

21:21:00 drmeister Maybe not - but I could add it.

21:28:20 Bike well, i can just use dtrace

21:40:32 drmeister Yes - you can.

21:57:57 drmeister That stuff I implemented for timing was pre-dtrace.

21:58:25 drmeister Note - I can only get dtrace to work with cclasp-boehm or iclasp-boehm when linked using 'obj' rather than 'thinLTO'

21:59:01 Bike really? i haven't noticed anything.

21:59:18 Bike you mean the image itself can't be LTO?

21:59:27 drmeister Really? dtrace would segfault for me.

21:59:55 drmeister I even posted on IRC/freenode #dtrace - they were very surprised.

22:02:24 Bike well, i think i'm using lto, anyway

22:28:14 drmeister Bike: Are you running in slime?

22:28:23 Bike yes.

22:29:14 drmeister Can you evaluate cmp::*dispatcher-count*

22:29:37 drmeister While or after it's finished loading that file?

22:29:48 drmeister That keeps track of the number of dispatchers that have been compiled.

22:30:14 drmeister https://github.com/drmeister/clasp/blob/cst/src/lisp/kernel/cmp/cmpfastgf.lsp#L855