03:15 colomon joined 03:58 vendethiel joined 04:02 FROGGS_ joined 04:25 vendethiel joined 05:39 vendethiel joined 06:26 vendethiel joined 06:51 vendethiel joined 08:08 lizmat joined 08:26 vendethiel joined 09:27 vendethiel joined, kjs_ joined 09:55 vendethiel joined 10:56 FROGGS[tab] joined 11:10 Ven joined 11:26 rurban_ joined 11:37 dalek joined 12:52 Ven joined 13:31 zakharyas joined 13:34 Ven joined 14:14 Ven joined 14:38 kjs_ joined 14:45 Util joined
dalek arVM/frame-gc-opts: 6adf362 | jnthn++ | src/core/frame. (2 files):
Add flag for when a frame references gen2 only.
 14:56
arVM/frame-gc-opts: 0428562 | jnthn++ | src/ (2 files):
Add write barrier macros for frames.
arVM/frame-gc-opts: ef48dd1 | jnthn++ | src/ (7 files):
Add frame barriers for various lexical writes.
arVM/frame-gc-opts: 11a54a5 | jnthn++ | src/gc/roots.c:
Initial, simple use of the frame gen 2 only flag.

This just uses it to elide scan_registers, which is the easiest thing we can do with it though with the least benefit. It will, however, let us see if there are any bugs with the frame write barrier stuff so far, and saves enough time to pay the cost of the barrier.
The next step will be to raise it to cover the MVMFrame overall; the end goal is to eliminate refs_frames on REPRs, and vastly reduce the size of the inter-generational set.		 15:01
jnthn If anybody would like to run a test on this, it'd be appreciated 15:03
You also need the moar-gen2-frame-opts branch 15:04
(in Rakudo)
It's not really an improvement *yet*, except we seem to pay off the cost of the new barriers.
We elide 6.1 million register scans during the Rakudo CORE.setting build 15:05
15:17 rurban_ joined
timotimo i'm glad this gets some attention 15:18
huh, what. when i panda rebootstrap i get write string requires an object with REPR MVMOSHandle 15:19
for pretty much every test
15:28 kjs_ joined 15:49 Ven joined
timotimo well, at least the module tests explode with MVM_SERIALIZATION_LAZY defined to be 0 16:16
so that's apparently not that?
jnthn timotimo: FROGGS++ did an ecosystem test with that in place 16:17
timotimo do we know what causes the problem? 16:18
jnthn I didn't know there was a problem until you mentioned it a moment ago 16:19
timotimo ah
jnthn I see the issue here also 16:21
timotimo well, better than "cannot write_string with REPR MVMOSHandle" would be :P 16:22
jnthn No idea what it si
*is
Yeah, that's what I get
timotimo i'm bisecting (more or less) rakudo right now 16:23
jnthn ok
It failed a clean bootstrap also
timotimo you mean a panda bootstrap? or a nqp bootstrap? 16:24
jnthn panda
timotimo yeah, that does fail
jnthn All NQP/Rakudo stuff looks fine 16:25
timotimo on my machine also, i mean
jnthn spectest run was good
timotimo correct
on edb5d9f panda is happy 16:26
interesting 16:27
your $*VM commit triggers the problem
jnthn wtf 16:28
timotimo yes, very
jnthn With the installed perl6-m I can do -e "use Test; plan 42" and it works just fine
timotimo it may be necessary to load another extra module 16:29
jnthn Adding a use NativeCall doesn't trigger it 16:31
uh
Yeah, doesn't. Either before or after the "use Test"
16:34 kjs_ joined
dalek MoarVM: 7955cee | fperrad++ | src/moar.h: 16:40
MoarVM: fix for ARMv5
MoarVM:
MoarVM: When cross compiling for ARMv6, all works fine with `libatomic_ops`.
MoarVM:
MoarVM: For ARMv5, the linking fails with :
MoarVM: ```
MoarVM: linking moar
MoarVM: arm-none-linux-gnueabi-gcc -o moar -O2 -DNDEBUG -Wl,-rpath,/usr/lib src/main.o -L. -lmoar
MoarVM: ./libmoar.so: undefined reference to `AO_compare_and_swap_full'
MoarVM: ./libmoar.so: undefined reference to `AO_fetch_compare_and_swap_full'
MoarVM: ./libmoar.so: undefined reference to `AO_fetch_and_sub1_full'
16:41 colomon joined
dalek MoarVM: bba0d8c | (Francois Perrad)++ | Configure.pl: 17:33
MoarVM: Configure: add an option --ar
MoarVM:
MoarVM: currently, the option --cc and --ld are available, but it is not enough when cross-compiling.
MoarVM:
MoarVM: the current workaround is:
MoarVM: ```
MoarVM: $ make AR=<ar>
MoarVM: ```
MoarVM: but the right way must be:
MoarVM: ```
MoarVM: $ perl Configure --ar <ar>
17:47 colomon joined 17:51 kjs_ joined 18:01 kjs_ joined
timotimo that's quite the commit message! 18:11
tadzik oh! 18:14
that might help Perl 6 on Sailfish
18:21 kjs_ joined 18:27 colomon joined 18:31 camelia joined 18:41 colomon joined
FROGGS_ Arrr! 18:48
x.o 18:49
18:49 camelia joined
timotimo :) 18:50
18:52 camelia joined 18:54 Ven joined, camelia joined 18:56 camelia joined 18:59 camelia joined
TimToady the 👀 have it 19:00
timotimo hark! 19:21
19:31 Ven joined
timotimo would people be mad if i wrote some code in the jit compiler to get better code output for the common tc, st, object, objectbody arguments that start pretty much every reprop? 21:05
jnthn Well, depends if it's nice correct code, or nasty busted code... :P 21:07
Geez MSVC, how many errors can you produce from a one-char typo... 21:08
timotimo i'd just like to have fewer movs 21:11
dalek arVM/frame-gc-opts: 64ac9a7 | jnthn++ | src/ (2 files):
Bring dynlex_cache_name under gen2 frame scheme.
arVM/frame-gc-opts: 09c41d7 | jnthn++ | src/gc/roots.c:
Lift "did we see gen2" code higher.

In preparation for brining more things under the generational frame scheme.
jnthn timotimo: Yeah, I agree that's a chance to do a bunch better there 21:12
timotimo: brrt may feel it will make the exprcomp stuff harder, but I'm not sure it will be a big issue
timotimo i suppose we'll have to wait a bit for the exprcomp stuff
jnthn Well, that'll be the clean way to do it 21:13
Anything done now will be a sort of "hack" in a sense
timotimo i'd tend towards putting in something that'd be very easy to remove again
agreed
hmm 21:16
nah, i'll leave that for later
great, nqp building says the next op to implement is setelemspos 21:18
very few arguments, low surprise factor
dalek arVM/frame-gc-opts: d07416b | jnthn++ | src/gc/roots.c:
Special return data mark not needed on gen2 frames

It is only relevant for those on the call stack, and so already gone by the time something is marked as a gen2 only frame.		 21:22
arVM/frame-gc-opts: c29d386 | jnthn++ | src/gc/roots.c:
Continuation tags are ignorable on gen2 frames.

These are cleared up as a frame is removed from the call stack, which is a pre-requisite for being a gen2-only frame anyway.
timotimo 1098 devirt: emitted a setelemspos via jgb_consume_reprop
45 devirt: please implement emitting repr op deletekey 21:23
worth it? :P
414 devirt: please implement emitting repr op existskey
^- i'll do this one next
but first a rakudo build to gather more datas
dalek arVM/jit_devirtualize_reprops_2: 6d9773d | timotimo++ | src/jit/graph.c:
devirtualize setelemspos
 21:24
timotimo huh, how come existskey is :pure, but existspos isn't? 21:27
seems like a simple oversight
jnthn I think so
dalek arVM/frame-gc-opts: db14cae | jnthn++ | src/gc/roots.c:
Move code ref/static frame mark under gen2 scheme.

Since these are always set at the point of frame creation, no further hanlding is needed.		 21:30
timotimo it seems like we're already putting a whole bunch less things in the inter-generational set now? 21:32
i'd like to help figure out why our push_i (likely in cstack or bstack) isn't getting devirt'd 21:35
would that just be looking at spesh log output?
well, i suppose i ought to start with the source code
jnthn timotimo: Yes, I'd look at spesh log 21:36
No, we didn't reduce the size of the inter-gen set yet.
We are doing less work marking frames the only hang around to be closure outers for a long time, though. 21:37
This is a precursor to reducing the size of the inter-gen set.
(Like, this bit needs to come first.)
It's decidedly a case of "the first 80%" also.
timotimo oh, OK 21:38
famous 80%
jnthn I still need to think a little about the second 80%. 21:39
dalek arVM/jit_devirtualize_reprops_2: 6ffdced | timotimo++ | src/jit/graph.c:
devirtualize existspos and existskey
jnthn (To find the right factoring for it.) 21:40
dalek arVM: b5d57ae | timotimo++ | src/core/oplist:
mark existspos as pure
arVM/frame-gc-opts: 8bdff6f | jnthn++ | src/ (2 files):
Bring context object under gen2 frame scheme.

This one is only ever written in one location in the codebase, so the barriering bit is easy for it.		 21:42
jnthn Anyway, there ends the first 80%. 21:43
timotimo now to look at unshift and push again ... with hawk eyes
yo jnthn, do we ever remove facts if we've decided the guard that guarantees them has gone unneeded? 21:46
because i'm afraid relying on types in the jit won't be able to mark a guard as needed because spesh has already thrown out guards that have been deemed unneeded
jnthn timotimo: um...yeah, phase order strikes again 21:49
timotimo yup, always does
jnthn But yeah, you're right to be afraid. 21:50
Which may explain some of the weird errors you've seen.
What I suggest is
in optimize.c, spot the ops you can de-virt, check if they have a known type, and if so mark the guard as used
And leave a comment there saying "the JIT actually de-virts" 21:51
timotimo ah
yes, that's a much better design than what i'd've done intuitively right now
jnthn phew! Glad I was here to answer, then :) 21:54
timotimo we have way, way fewer repr ops that dispatch to optimize_repr_op in spesh than i've put into the consume_repr_op bit in the jit
but since that's just to dispatch to a repr's spesh method, it's not that bad
i must admit, i feel like i'm constantly and consistently bad at doing this sort of stuff ... 21:57
21:58 travis-ci joined
travis-ci MoarVM build failed. Timo Paulssen 'mark existspos as pure' 21:58
travis-ci.org/MoarVM/MoarVM/builds/55324483 github.com/MoarVM/MoarVM/compare/5...d57aee3691
21:58 travis-ci left
timotimo huh 21:58
also, in spesh bindattr and bindattrs were marked to have their "type" field as the 0th operand, isn't that wrong? 21:59
bindattr looks like bindattr($target, $CLASS, $attr, $value), no?
jnthn yes...where do you mean? 22:00
timotimo spesh's optimize.c; the "optimize_repr_op" has a "type operand" argument 22:01
that decides which operand's facts to check for "known type"
jnthn Oh
But you spesh the bindattr by the type of the object you have
timotimo oh
in that case, my jit implementation is wrong
jnthn Since it's REPR(target)->bindattr in the normal impl 22:02
uh, or whatever the function is actually called
timotimo yeah 22:03
but can you actually have a different REPR for a class and its superclass? 22:05
jnthn yes
It's orthogonal.
timotimo i mean, we put attributes inherited by a role into the class as visible to the class, so role vs class doesn't apply
OK
jnthn m: say array ~~ Any 22:06
camelia rakudo-moar 0e5632: OUTPUT«True␤»
jnthn m: say array[int8].REPR; say Any.REPR;
camelia rakudo-moar 0e5632: OUTPUT«VMArray␤P6opaque␤»
timotimo oh, good point
jnthn Yes, it'd be weird on the attribute case, but in general - it comes up.
timotimo and now to find out if "merge_facts_at_phi" collaborates well with this branch 22:08
fish: Job 1, “./perl6-m --target=mbc --output=lib/Test.pm.moarvm lib/Test.pm” terminated by signal SIGSEGV (Address boundary error) 22:13
;(
really. it works fine when run under gdb? 22:17
jnthn vasanalgrind to the rescue! 22:19
timotimo simply gave emit_call_c 6 args, but args[] only had 5 entries
dalek arVM/jit_devirtualize_reprops_2: 5d4fcdc | timotimo++ | src/jit/graph.c:
bindattr/getattr decide their REPR on their object operand

not the class operand!		 22:22
arVM/jit_devirtualize_reprops_2: b3f729b | timotimo++ | src/spesh/optimize.c:
spesh's optimize_repr_op must mark type guards used
arVM/jit_devirtualize_reprops_2: 698efb7 | timotimo++ | src/jit/graph.c:
fix number of arguments for existspos/existskey
timotimo weird, with merge_facts_at_phi i get a few more cases of "type unknown" :\ 22:30
but the number of "emitted" reprop devirts changes by only a tiny amount 22:32
22:33 ggoebel joined
timotimo how often do we do !cursor_capture? 22:44
because we push_i into $!bstack four times, which translates to four getattr invocations 22:45
which translates to four sp_p6ogetvt_o with a spesh slot each 22:46
doesn't solve the "unknown type" problem of course
doesn't seem to influence stage parse either 22:48
in nqp, is my $nfa := self.HOW.cache_get(self, $name); 22:49
any slower than my $nfa := self.^cache_get(self, $name); would be?
nqp: say NQPMu.^name
without the self, of course
nqp-m: say(NQPMu.^name) 22:50
22:51 camelia joined
camelia nqp-parrot: OUTPUT«Can't exec "./rakudo-inst/bin/nqp-p": No such file or directory at lib/EvalbotExecuter.pm line 188.␤exec (./rakudo-inst/bin/nqp-p /tmp/tmpfile) failed: No such file or directory␤Server error occurred! Closing Link: ns1.niner.name (Quit: camelia)␤Lost connect…» 22:52
..nqp-moarvm: OUTPUT«Confused at line 2, near "say NQPMu."␤ at gen/moar/stage2/NQPHLL.nqp:493 (/home/camelia/rakudo-inst-2/share/nqp/lib/NQPHLL.moarvm:panic:105)␤ from gen/moar/stage2/NQP.nqp:921 (/home/camelia/rakudo-inst-2/share/nqp/lib/nqp.moarvm:comp_unit:872)␤ from gen/…»
..nqp-jvm: OUTPUT«(signal ABRT)#␤# There is insufficient memory for the Java Runtime Environment to continue.␤# pthread_getattr_np␤*** Error in `java': double free or corruption (!prev): 0x00007f001001a8c0 ***␤# An error report file with more information is saved as:␤# /tmp/…»
timotimo also, we look up $?CLASS like four times as well
many times in Cursor.nqp, i bet 22:57
i don't know how expensive getlexperinvtype is, though
jnthn getlexperinvtype can spesh into a constant 22:59
timotimo ah, that's good to know 23:00
let me see if it does
jnthn About self lookup - in many cases it is reduced to a register access
(by lex -> loc lowering the NQP compiler)
Even if not, lexical access is cheap. 23:01
timotimo OK, good to know
in this instance, it seemingly turns into a getspeshslot once and another time it doesn't 23:02
possibly because of the branch that checks if $!cstack is defined; possibly doesn't get run often enough to get good log data
seems like all the !cursor_* methods are very cheap, all in all 23:05
jnthn good to know :) 23:11
Gonna go rest/recover more... 'night 23:12
timotimo good night!
recover well!

Powered by Cro and the Raku® Programming Language.

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

Thank you!