00:28 Kaiepi joined 01:38 brrt joined 02:37 brrt joined 02:56 ilbot3 joined 03:26 lizmat joined
nine MasterDuke: that must be an issue. One gets away with a lot of inaccuracies in C because the compiler does implicit casts. Not so much in assembly. For a quick tests you could change MVM_spesh_log_is_logging's signature to return an MVMint64 09:41
MasterDuke: better in the long term would be to just inline the function (as it's meant to be according to the comment above it). It's just: tc->cur_frame->spesh_cand == NULL && tc->cur_frame->spesh_correlation_id && tc->spesh_log 09:42
11:52 brrt joined
brrt good * 11:57
nine brrt: whatever good to you, too! 12:10
brrt: irclog.perlgeek.de/moarvm/2018-02-17#i_15828814 don't you mean some 0 value that is smaller than 64 bits?
MVM_spesh_is_logging returns an MVMint32, but I guess the code the JIT creates for an NZ test looks at the full 64 bit register? 12:11
brrt correct 12:12
but that can not really be the problem
as I understand the x86-64 ABI, an assignment of 0 to a 32-bit register, or indeed, the assignment of anything to a 32 bit register
will automatically clear the upper 32 bits 12:13
nine Even return values of functions?
brrt yes
smaller values, not
because the return value is stored in %eax (or %rax)
so, hypothesis disproved, i say
:-) 12:14
nine We could easily test the hypothesis by changing MVM_spesh_log_is_logging's signature to return an MVMint64
brrt true 12:15
but i'm not getting a replication right now, on macOS + clang 12:16
I did get them when running with --debug --optimize=0
so there's something going on there, too
bbiab 12:22
jnthn Wait, are we calling MVM_spesh_log_is_logging from the JIT? 12:50
That should never happen
Logging is something the interpreter does to figure out what types are showing up
Once we're in specialized code, we're already optimizeed 12:51
*optimized
And there's always a ->spesh_cand
timotimo i told them that :| 12:57
13:08 brrt joined
brrt not only that, we're calling MVM_spesh_log_is_static, from the JIT 13:09
yeah, i'm puzzled as well
i mean,
i can fix it with not emitting that code
but why is it returning true in the first place?
MasterDuke fwiw, i did try changing MVM_spesh_log_is_logging signature to an MVMint64 and still got segfaults 13:11
jnthn Dunno, maybe that means there's some other bug
Is tc getting passed correctly? 13:12
But yeah, we should remove any calls to that
MasterDuke tc=0x555555758c40 13:13
brrt print *tc 13:14
wil give information
that seems high though
usually tc is allocated to a lowish page
MasterDuke i'll trying rebuilding and rerunning gdb and see what that says 13:15
gist.github.com/MasterDuke17/9f386...5fbab53376 updated 13:16
jnthn Hm, looks legit at least 13:18
oh, though if it's about the is_logging returning true when it shouldn't, then it'd be the thing we passed there that's in question
Hm, I thought the "are we spesh logging" thing was a static inline function 13:19
brrt i'm thinking of a MVM_spesh_debug_is_active() to replace all checks for tc->instance->spesh_log_fh
compilers will create an uninlined variant of that if you take a pointer
anyway
MasterDuke gist updated with print *value output
jnthn Ah, ok
brrt one think you can do is wrap the (when) with a (guard (when ...) control_breakpoint 0) 13:20
then under a debugger (*not* gdb) you will be interrupted and can step through
MasterDuke lldb? 13:21
brrt (gdb will assume it has placed the breakpoint, will try to restore an instruction, presumably from a hash table, will reinsert the an instruction from uninitialized memory, and crash)
lldb will work
i mean to fix this in gdb at some point :-) 13:22
seems like a legit *tc
yes
the other hack i sometimes do, is to copy or wrap the definition of the called function, e.g. MVM_spesh_is_logging, and replace it with MVM_spesh_is_logging_jit_version() 13:23
and then break on that
MasterDuke gotta go feed a baby, might be back in a bit 13:24
brrt :-)
timotimo: you had a tool to visualize spesh logs, didn't you 13:26
can i ask why there's a file called 'gh-pages-deploy_key.enc' in our 'tools' directory 13:31
hmmm, graph_spesh.p6 seems to have bitrotted 13:33
there's no MAST::Ops
can we still load a '.moarvm' file as a library 13:36
aha, i have a hypothesis 13:43
i'm seeing a 'lowered' istrue+unless_i, where the frame-handler (start, goto) annotations are on the unless_o, and the istrue on the unless_i 13:44
MasterDuke i'm in lldb, anything i can do in ~3m?
brrt so those need to be moved to the istrue
i dunno
can you break set --name MVM_spesh_is_logging 13:45
well, better yet
MasterDuke i did the guard thing
brrt oh
can you run it
(it'll break before you enter the MVM_spesh_is_logging)
then, when you have, use 'step' ('s') or 'step-instruction' ('si') to step into the frame 13:46
MasterDuke gist updated
brrt that will be a call
(callq)
ok, step until you return from the frame, and report the result of 'print $rax'
MasterDuke gist updated, was that what you meant? 13:47
brrt continue pls :-) 13:48
MasterDuke or step further?
brrt step further, and print all you can :-)
the $rax is the value after the 'return'
oh, and can you print *cur_frame 13:49
MasterDuke updated 13:50
or did you want cur_frame inside MVM_spesh_is_logging? 13:53
brrt kind of, but never mind :-) 13:56
hmm 13:57
see, $rax is 0
ok, can you step further and see if you enter MVM_spesh_log_static
maybe it doesn't happen in the first run
MasterDuke yep
took about 8 steps 13:58
brrt (i sometimes use the tc->frame_sequence_nr for determining the correct frame to break in)
uhuh
MasterDuke afk for a while, but i'll leave lldb running if there's anything i can do when i get back 14:00
brrt awesome
(i hope to replicate it myself by that time, but alright) 14:01
I think I have a fix for that issue
timotimo brrt: that module exists, but it's under tools/lib or something 14:04
brrt uhuh
I think I have a fix for zoffix's issues 14:09
then again... 14:11
.ask jnthn I think when I split the if_o into istrue + if_i, I think i need to move at least some annotations onto the istrue, but possibly not all 14:12
yoleaux brrt: I'll pass your message to jnthn.
brrt I'm not sure which is which, to be honest
timotimo mhh, good idea
well, istrue is the one that might invoke, so it probably wants most annotations
brrt uhuh
well, some instructions are supposed to be 'before' an instruction, and some are 'after' 14:14
timotimo s:1st/instructions/annotations/ 14:15
brrt right :-)
timotimo oh, so after splitting, we do need to put those that were meant to come after on the second resulting instruction 14:18
here's a thought that makes things much easier:
we introduce another instruction that is exactly like if_o and unless_o, but it takes a third argument, which is just a scratch register that spesh allocates for us 14:19
brrt uhuh....
i'm listening
i'm skeptical but listening :-)
timotimo that's already everything
brrt oh,
i see
yeah, and then do the split just in the JIT 14:20
hmmm
timotimo we'll no longer need to split up the instruction, because we now have the register "for free"
aye, jit is already rather "splitty" by its nature
brrt nah, i think this is quite alright
timotimo also, we won't be able to forget adapting the splitting code to new annotations as they show up in the future
the interp.c implementation of sp_if_o (or what we end up calling it) will be identical to the if_o and unless_o ones we had so far 14:22
the only effect on speshed-but-not-jitted will be that the register set will grow a little
brrt uhuh 14:29
then again, i never really quite liked the implementation in jit/graph.c 14:30
timotimo is going to do some grant work now 14:34
i think next up is making gc recording "proper" when multi-threaded 14:44
Geth MoarVM: 22d2db5e08 | (Bart Wiegmans)++ | 2 files
[Spesh] Move annotations to istrue when split from if_o

In commit c7774e37d2 we unconditionally split if_o/unless_o into istrue + if_i/unless_i, which fixed rakudo issue #1483.
However, the annotations that would indicate the start of the if_o instruction, would now inadvertently be placed in the middle of the ... (8 more lines)		 14:48
timotimo re-fiddled the datastructure for profiles a little, and also how the sql output backend goes through the threads object 15:22
now the threads object should have all threads in it, also the "profile" table will have one entry per thread, including its thread id
15:27 zakharyas joined
nine JFTR spectest seems to be back to the usual 2 backtrace related failurs with BLOCKING and NODELAY except for t/spec/S09-typed-arrays/hashes.t which fails with "Incompatible MROs in P6opaque rebless for types NQPClassHOW and QAST::WVal+{QAST::SpecialArg}" 15:29
.tell brrt MVM_SPESH_BLOCKING=1 MVM_SPESH_NODELAY=1 perl6 -Ilib t/spec/S09-typed-arrays/hashes.t 15:30
yoleaux nine: I'll pass your message to brrt.
nine .tell fails with "Incompatible MROs in P6opaque rebless for types NQPClassHOW and QAST::WVal+{QAST::SpecialArg}" but passes with MVM_JIT_EXPR_DISABLE=1
yoleaux nine: I'll pass your message to fails.
nine .tell brrt fails with "Incompatible MROs in P6opaque rebless for types NQPClassHOW and QAST::WVal+{QAST::SpecialArg}" but passes with MVM_JIT_EXPR_DISABLE=1
yoleaux nine: I'll pass your message to brrt.
15:30 zakharyas joined
timotimo pro-tip: make sure you configure with the right --prefix... 15:32
nine timotimo: oh yes, that one can bite sooo hard 15:34
15:34 zakharyas joined 15:37 zakharyas joined 15:42 zakharyas joined
timotimo i was ack-ing the source for the message that was messing up my display, but i couldn't find it :D 15:45
Unhandled exception: const_n32 NYI - damn it! 15:53
nine .tell brrt Please disregard that. I didn't realize that I was running with an experimental patch applied. 15:59
yoleaux nine: I'll pass your message to brrt.
timotimo oh, fun! joining a thread destroys the tc, and with it all profiler-related data 16:18
16:34 robertle joined 16:38 AlexDaniel joined
MasterDuke timotimo++ --profiling multi-threaded will be awesome 16:38
timotimo next up will be figuring out what causes the start time of a node to not be set 16:40
INSERT INTO profile VALUES (18446635666303613,0,1);
^- this doesn't make much sense :)
MasterDuke fwiw, i still get the segfault when building nqp if i enable getlextatic_o on top of HEAD 16:45
timotimo the problem was a problem i thought i'd fixed before o_O 16:51
anyway, that's kinda nice now
MasterDuke are these mostly moarvm fixes or nqp? 16:55
timotimo changes and fixes in moarvm, changes to fix stuff and changes to work with the new data in nqp 16:57
rsi rest break time
though actually, let me push what i have so you can play around with it for a moment :) 16:58
Geth MoarVM/profiler_new_spesh_semantics: 8 commits pushed by (Timo Paulssen)++
timotimo the branch name is a bit strange, because the first thing was to make the profiler UI not show spesh time as time "not spent running code" since it moved to another thread 16:59
but really i'll want to teach the profiler about vars like SPESH_BLOCKING so that can be actually correct
MasterDuke cool, i'll play around with it 17:03
17:05 Ulti joined
Ulti just an FYI, this dir is mostly full of 0byte files so I assume something somewhere is a bit wonky www.moarvm.org/measurements/perl6-bench/ 17:05
\o.
timotimo MasterDuke: the html frontend doesn't know what's up with extra threads yet i believe 17:10
so only the sql makes much sense right now
MasterDuke looks like you made some changes to the html though?
timotimo barely
MasterDuke ah 17:11
timotimo in the future our profiler frontends will have to merge trees from different threads, i assume 17:15
Geth MoarVM: jstuder-gh++ created pull request #806:
Rakudo RT#102994: Add flag indicating HLL init status on CodeRef (CodeRef edition)
synopsebot RT#102994 [open]: rt.perl.org/Ticket/Display.html?id=102994 [BUG] State variables are never initialized if the first call of the containing block does not reach them
17:25 zakharyas joined
timotimo MasterDuke: looks like i forgot to commit syntax/typo fixes for the generated sqlite code 17:29
AlexDaniel Ulti: heh! What a find 17:31
timotimo neat. p6profiler-tools's graphviz target already shows every thread entry
yeah, jnthn's home desktop (i think?) used to generate those and one day it just stopped generating actual results 17:32
Ulti AlexDaniel: yeah I noticed at Xmas for the blog post about perf but didnt mention it
AlexDaniel I didn't know this page existed :)
Ulti then recently came up in my browser autocomplete
AlexDaniel (browser autocomplete)++ 17:33
timotimo oh, maybe i just forgot "make install"
AlexDaniel samcv: I think we'll be ready much sooner this time. How is the changelog going? 17:42
17:56 domidumont joined
MasterDuke timotimo: i attempted to --profile `my @p = (^30_000).race.grep(*.is-prime); say @p.elems`, but it seemed to hang after `took data from main thread` at 100% cpu (i killed it after a couple min) 18:04
timotimo oh, that must be the other thing i fixed and apparently didn't commit 18:05
it's probably entered gc
oooh 18:06
Geth MoarVM/profiler_new_spesh_semantics: c6d1115d7a | (Timo Paulssen)++ | 3 files
record a gc's sequence number
MoarVM/profiler_new_spesh_semantics: b6e8639302 | (Timo Paulssen)++ | src/profiler/instrument.c
keep main thread in gen2 mode

as the other functions use its tc to allocate objects.
timotimo try this
oh whoops? 18:09
MasterDuke `MoarVM oops: Cannot leave gen2 allocation without entering it`
timotimo yeah 18:11
Geth MoarVM/profiler_new_spesh_semantics: 3484ba9189 | (Timo Paulssen)++ | src/profiler/instrument.c
keep main thread in gen2 mode

as the other functions use its tc to allocate objects.		 18:12
timotimo ^- a force-push
MasterDuke btw, do you plans include incorporating --telemeh output in profiles?
timotimo not sure yet 18:13
MasterDuke ooo, the --profile succeeded!
timotimo i don't have a good reason for it yet, but i do have more timeline-like data now
like the absolute time a gc started will be interesting i bet
RSI rest break & 18:14
MasterDuke heh, 10m file though, not gonna open that in my browser
1524ms in Channel.pm's receive for `my @p = (^10_000).race.grep(*.is-prime); say @p.elems`, and it's not jitted 18:17
timotimo that's most likely almost exclusively waiting for something to become available on the channel 18:24
m: say (^10_000).pick(*)[*-1]; say now - INIT now 18:25
camelia 6433
0.03790328
timotimo i wonder if shuffling the input numbers would do you much good, because lower numbers are quicker to test (i guess at least?)
AlexDaniel 99999999999999999999 .is-prime.say 18:32
evalable6 False
timotimo scratch that, it most probably takes about the same amount of time 18:37
unless you reach big integer territory
hm, though actually, i bet we force a big int for .is-prime even if we fit into 32bit
19:00 zakharyas joined 19:02 buggable joined 19:23 Geth joined
timotimo it should make sense to take the lowest start time for a given gc sequence number and the maximum of start time + time to figure out how long a given gc run took in total, right? 19:27
because the program resumes only when all gcs have finished?
occasionally i still get "profiler lost sequence"; i'd really like to know how that happens ... 19:33
19:57 Geth joined 20:06 Geth joined 20:08 zakharyas joined
MasterDuke timotimo: you think it would make sense to have an is_prime in addition to is_prime_I? or MVM_int_is_prime in addition to MVM_bigint_is_prime? 20:42
but yeah, MVM_bigint_is_prime does a `mp_int *ia = force_bigint(ba, tmp);` 20:43
timotimo i don't think it should be an extra op; it should probably just decide based on the size 21:02
21:49 Kaiepi joined 21:58 zakharyas joined
timotimo now i have to decide whether i want to accept inclusive times > 100% 22:17
tbh, i could go that route 22:18
fun! i was wondering what in the heck was calling "sleep" so much in my profile. turns out i get the thread pool scheduler's supervisor thread in my profile now <3 22:24
currently we're removing retained_bytes + gc_promoted_bytes from cleared_bytes to get the final cleared_bytes count 22:43
it turns out, however, that gc_promoted_bytes also contains the unmanaged size of objects
a few times in this particular profile we ended up promoting more bytes than we freed, thus causing the cleared bytes count to go negative
i might want to split up those numbers into bytes "directly" promoted from the nursery and unmanaged bytes promoted out of the nursery 22:44
it's especially muddy because the unmanaged size of objects can change over their lifetime, whereas the regular size will not (except when we mixin? but i think that has a pointer and adds to unmanaged size instead?) 22:45
m: say 140106937758704.base(16) 22:48
camelia 7F6D304173F0
timotimo i don't know how we reach this kind of size, though
OK, so there's a few SpeshLogs which are all 384 KBytes, 1397 KBytes in one SCRef, 1237 KBytes in an NFAType 22:53
still don't know how it reaches that incredible number, though 22:54
m: say 34677491 / 140106937758704 22:57
camelia 0.0000002475073080
timotimo i can account for this tiny amount over the whole run time of the program, but the huge number was from just a single gc run
rsi rest break & 22:58
AlexDaniel timotimo: by the way, that nice piece of software for taking breaks… 23:04
timotimo: had to disable it because it was getting too much in the way :(
but the concept is wonderful and the execution is great too, just doesn't work for me I guess 23:05
( www.workrave.org/ )
timotimo have you tried its less intrusive mode? 23:09
"quiet" mode

Powered by Cro and the Raku® Programming Language.

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

Thank you!