github.com/moarvm/moarvm | IRC logs at colabti.org/irclogger/irclogger_logs/moarvm
Set by AlexDaniel on 12 June 2018.
01:03 sena_kun joined 01:05 Altai-man_ left 01:10 Kaiepi joined 01:27 AlexDaniel joined, AlexDaniel left, AlexDaniel joined 01:52 farcas1982regreg joined
[Coke] m: use nqp; dd nqp::hllboolfor(1,"boom") 02:03
camelia (signal SEGV)
[Coke] ^^
bug nqp#619
linkable6 NQP#619 [open]: github.com/Raku/nqp/issues/619 hllboolfor op SEGVs
03:02 Altai-man_ joined 03:05 sena_kun left 05:03 sena_kun joined 05:04 [Coke] joined, [Coke] left, [Coke] joined 05:05 Altai-man_ left 07:02 Altai-man_ joined 07:05 sena_kun left
MasterDuke jnthn: github.com/Raku/nqp/issues/619 is actually spesh 07:38
08:27 zakharyas joined 09:03 sena_kun joined 09:05 Altai-man_ left 09:20 MasterDuke left 09:30 zakharyas left 09:32 zakharyas joined 09:42 zakharyas1 joined 09:44 zakharyas left 10:24 squashable6 left 10:25 squashable6 joined 10:43 zakharyas1 left 11:02 Altai-man_ joined 11:05 sena_kun left
nwc10 jnthn: I hope that your beer fridge operative is doing better than mine. I went to get a beer, and there were none of the one that I wante *in* the fridge. (Almost) two crates worth next to it, but none inside. I had to settle for a substitute. 11:17
can't get the staff :-(
lizmat feels for nwc10 11:24
timotimo hey nwc10 11:32
imagine >>+<< and friends, as well as >>.blah and -<< were many times faster than they currently are (for native arrays)
got any ideas what you'd implement with that? 11:33
nwc10 I admit that I can't remember what the syntax means, but the answer I'd like to be able to give is "character classes in regular expressions". But I am unfamilar with *anything*/everything, apart from "OMG, there are too many characters in Unicode to keep implementors sane" 11:35
timotimo ha 11:36
those are hyper-operators
like elementwise addition of vectors is >>+<<
they behave much like Z, but >>+<< requires both sides to be the same length, rather than just stopping at the shorter one 11:37
and also they are eager rather than lazy
jnthn nwc10: While operations are going OK, procurement might need to get its finger out soon...stocks are dwindling, especially of imperial stouts/porters. 11:41
nwc10 :-) or should that be :-( 11:42
jnthn Situation is not critical yet )
nwc10 that is good to hear.
dogbert17 in case anyone is bored and wants to look at gdb output instead I added some to github.com/Raku/nqp/issues/619 11:52
nine would love to be bored some time... 11:53
timotimo is making hypers run at hyperspeed
dogbert17 so no one is hyperbored then :)
timotimo hyperborea? 11:55
nine my guess is actually correct 11:56
dogbert17 you're good at guessing, for me it crashes even if spesh is disabled 11:57
timotimo yeah, hllboolfor doesn't check the return value of get_hll or whatever 11:59
question is what it should return for a bogus hll
making it throw an exception means to also set :throwish on the op
nine Apparently MVM_hll_get_config_for will happily return an empty config for non-existing hlls
timotimo which has effects on spesh and jit
nine But none of the ops that throw exceptions on invalid use are marked :throwish 12:02
12:05 MasterDuke joined
timotimo that .. might be bad? brrt would know 12:05
when perf says this code spends 34% of its time in sin_avx, i think i can be happy: 12:08
m: my int @a = ^100_000; my int @b = ^100_000; my int @c = ^100_000; my int @d = ^100_000; my int @result = ^100_000; my $now = now; for ^10 { @result = @a >>+<< @b >>+<< @c >>+<< @d>>.sin; Nil }; say now - $now;
camelia This type cannot unbox to a native integer: P6opaque, Num
in block <unit> at <tmp> line 1
timotimo ah, right
m: my int @a = ^100_000; my int @b = ^100_000; my int @c = ^100_000; my int @d = ^100_000; my num @result = ^100_000; my $now = now; for ^10 { @result = @a >>+<< @b >>+<< @c >>+<< @d>>.sin; Nil }; say now - $now;
camelia This type cannot unbox to a native number: P6opaque, Int
in block <unit> at <tmp> line 1
timotimo mhh?
anyway, my impl doesn't bork on this, which is actually bad because it doesn't always kick in, and if a slight change by a user takes their code from my impl to the regular one, getting type errors like this could be very confusing 12:09
12:44 zakharyas joined 12:50 sena_kun joined 12:51 Altai-man_ left 14:00 zakharyas1 joined 14:02 zakharyas left 14:17 MasterDuke left 14:22 AlexDaniel left 14:35 MasterDuke joined 14:37 zakharyas1 left 14:46 zakharyas joined 14:49 Altai-man_ joined 14:51 sena_kun left 15:04 lucasb joined 16:07 AlexDaniel joined, AlexDaniel left, AlexDaniel joined
nine I dare say there's something to it: Missing root for `thread_obj` in MVM_gc_mark_thread_blocked (child_tc); at src/core/threads.c:126 used in thread_obj.15_22 = thread_obj; at src/core/threads.c:147 16:17
Same goes for Missing root for `config_hash` in MVM_intcache_for (tc, _143); at src/core/hll.c:157 used in D.18030 = config_hash; at src/core/hll.c:159 16:20
Also several of these: Missing root for `t` in msg_box = MVM_repr_box_str (tc, _45, msg_str); at src/io/asyncsocket.c:127 used in _48 = t->body.queue; at src/io/asyncsocket.c:127 16:22
Geth MoarVM/gcc_root_checker_plugin: 57e798527d | (Stefan Seifert)++ | tools/check-roots.py
Add a full list of MVMCollectable types to check-roots.py
 16:23
nine Time to throw away the whole algorithm 16:31
timotimo oh?
start afresh with a different approach now that you've learned something?
nine Yes. I started out with going through basic blocks and instructions (nice to see the same terminology used) and taking note of things, i.e. collect some state and when I encounter a variable look at the state. 16:32
That breaks down hard in the face of branching and loops. 16:33
So instead, I'll now work my way backwards from every read of a variable through the control flow graph, take every branch and make sure that I will go through each loop at least once. 16:34
Should be relatively simple by just going through each BB's preds and add them to my array of seen blocks only after I followed that link. 16:35
Only thing I don't yet know: I'll have to take each of these possible paths through the control flow and process them individually, i.e. simulate the temp root stack for each path. Don't know how to bring that into my loops/recursions 16:40
16:50 sena_kun joined 16:51 Altai-man_ left 16:55 zakharyas left 17:08 zakharyas joined
timotimo store what state your stuff has when it reaches the start of a BB and only consider a link "seen" if both the target BB (source, really) and "state at that position" were seen as a combination 17:11
17:36 zakharyas left
nine God damn it Python! 17:49
def collect_control_flows(bb, path=[], seen={})
When called with only one arg will always get the same [] and {}, so when you modify them, it carries over
moritz that's why you default to None 18:12
and have the clunky if path is None: path = [] branch for each argument
timotimo ah, yes
a common pitfall 18:13
moritz or path = path or []
nine On the bright side: it seems to work! Now correctly detects the error in: void loop_notok(MVMThreadContext *tc, MVMObject *param) { for(int i = 0; i < 2; i++) { MVM_is_null(tc, param); MVMObject *alloced = MVM_repr_alloc_init(tc, tc->instance->boot_types.BOOTInt); } } 18:14
18:49 Altai-man_ joined 18:51 sena_kun left 19:31 patrickb joined 19:34 patrickb left 19:41 zakharyas joined 19:48 farcas1982regreg left
nine timotimo: Missing root for `ex` in request_all_threads_suspend (tc, ctx, 0B); at src/debug/debugserver.c:563 used in _20 = allocate_handle (tc, ex); at src/debug/debugserver.c:575 19:57
timotimo ex isn't used any more after that though? 19:59
it's pushed into the handles vector where it'd be marked by the gc, at least that's how i want it to behave :D
nine ex may have been moved by the request_all_threads_suspend call 20:00
The message reads like "you need to add an MVMROOT in src/debug/debugserver.c:563 to fix the breakage in src/debug/debugserver.c:575" 20:03
timotimo oh i missed the bit in the middle 20:06
smort 20:07
nine The exact same error in error handling code is in procops.c, filewatchers.c, 4 times in asyncsocket.c and twice in asyncsocketudp.c 20:12
All false positives I've looked at so far are because objects were allocated in gen2 directly (mostly deserialization and profiler). 20:17
timotimo ha 20:19
nine I'm not so sure about the static_frame cases though. Since freshcoderef allocates one in the nursery the assumption in static frame handling code may actually be wrong 20:20
Geth MoarVM/gcc_root_checker_plugin: d4637c1a0a | (Stefan Seifert)++ | tools/check-roots.py
Ignore allocations before variable gets overwritten with function results
 20:26
MoarVM/gcc_root_checker_plugin: 7553099753 | (Stefan Seifert)++ | tools/check-roots.py
Properly model control flow in check-roots.py

For each variable usage, we now trace back the control flow to the start of the function, while ensuring we go through each loop at least once, to get the correct state of the temp root stack at that location. With this we handle branches and loops correctly.
20:49 squashable6 left 20:50 sena_kun joined, squashable6 joined 20:51 Altai-man_ left 21:13 zakharyas left 21:18 squashable6 left, squashable6 joined 21:20 squashable6 left 21:21 squashable6 joined
[Coke] can someone explain one of the bindposXXX opcodes to me so I can document them all? Guessing it's, given an array of the right dimension, set the value at an index-tuple? (why is it "bind"?) 22:07
e.g. MVM_repr_bind_pos_3d_i? (I know the i is specifying the value type...)
lizmat [Coke]: there are specific opcodes for 2- and 3- and 3+ dimensional native arrays 22:08
*shaped
they come in _i, _s and _n tastes 22:09
and "" taste for opaques 22:10
and atpos* associates
as to why it is "bind": in nqp, there is no assignment :-)
you either bind a value, or a container with a value 22:11
[Coke] is atpos the "get" to bind's "set" ? 22:12
lizmat yup 22:13
m: use nqp; my int @a = 1,2,3; dd nqp::atpos_i(@a,2)
camelia 3
[Coke] Danke
lizmat m: use nqp; my @a = 1,2,3; dd nqp::atpos(@a,2)
camelia This type (Array) does not support positional operations
in block <unit> at <tmp> line 1
lizmat m: use nqp; my @a = 1,2,3; dd nqp::atpos(nqp::getattr(@a,List,'$1reified'),2) # slightly more complex in HLL land 22:14
camelia P6opaque: no such attribute '$1reified' on type List in a Array when trying to get a value
in block <unit> at <tmp> line 1
lizmat m: use nqp; my @a = 1,2,3; dd nqp::atpos(nqp::getattr(@a,List,'$!reified'),2) # slightly more complex in HLL land
camelia Int @a = 3
lizmat sleep& 22:16
[Coke] greppable6: atposnd 22:31
greppable6 [Coke], Found nothing!
22:49 Altai-man_ joined 22:51 sena_kun left
timotimo multidim arrays are barely performant, maybe that's why they aren'- used often 23:33
but also people aren't supposed to use nqp ops in the ecosystem :D

Powered by Cro and the Raku® Programming Language.

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

Thank you!