00:03 dalek joined
jnthn sleep & 00:16
00:16 tadzik joined 00:24 colomon joined 01:47 ilbot3 joined
lizmat checks whether he changes broke parrot 07:35
*her 07:36
07:54 nikolos joined 08:15 nevalgen joined
jnthn Turns out it's the dead instruction removal that hoses the trig tests 09:54
argh 10:26
atan2_n w(num64) r(num64) w(num64) :pure
Of cource, that third arg should be an r(num64)
Slo it was deleting instructions assuming the instruction really did write to its 3rd reg arg. 10:27
And there I thought we had a really weird SSA construction bug 10:28
decont r12(2), r13(11)
unbox_n r15(1), r12(2)
atan2_n r14(2), r14(1), r15(2)
Note the r15(2) on the end there. :)
So of course the unbox_n and decont were considered dead. 10:29
dalek arVM/spesh: b77ca87 | jnthn++ | src/core/ext.c:
Avoid accidental "pure" marks on extops.
 10:32
arVM/spesh: 4096332 | jnthn++ | / (3 files):
Correct atan2 op signature.

We read, not write, the 3rd register argument.
arVM/spesh: 1769ad6 | jnthn++ | src/spesh/facts.c:
Fix usage tracking of inc/dec instructions.

They are the only ones that read/write, and so need special handling in the usage tracking.		 10:56
jnthn just got his lowest "build Rakudo" time :) 10:57
(as in, Rakudo's "make" phase) 10:58
OK, spesh is doing good in Rakudo spectest now. Let's try and keep it that way. :) 11:03
11:55 ilbot3 joined 12:35 ilbot3 joined
dalek arVM/spesh: f0d1d72 | jnthn++ | src/ (5 files):
Stub in spesh guards for specializing on arg type.

None are produced so far; this just prepares the storage of them and adds a first pass at the code for checking them.		 14:13
FROGGS morning 14:21
jnthn o/ FROGGS
15:01 zakharyas joined
dalek arVM/spesh: 0848295 | jnthn++ | src/spesh/ (3 files):
Start adding guards and specializing by arg type.

Basically works, but we don't de-spesh yet on mixins, meaning we get some dispatches wrong and break some NQP tests.		 15:10
arVM/spesh: 6e04b28 | jnthn++ | src/spesh/candidate.c:
Fix inter-gen issue in spech candidate storage.
arVM/spesh: e3c17bc | jnthn++ | / (5 files):
Start labeling opts that are de-opt points.
15:13 colomon joined
JimmyZ Does it possible snapshot the optimized Bytecode and serialize it? 15:13
jnthn No. 15:14
Well, maybe, but it's highly non-trivial.
You'd have to re-construct all the safety proofs.
And some of them will depend on VM internals.
JimmyZ Yeah
jnthn So you'd end up with incompats. 15:15
So, quite nasty to do.
JimmyZ Or part of it, such as decont removing, set removing etc 15:16
What is Spesh short for? 15:22
FROGGS specialization 15:23
jnthn The "c" in "specialization" is pronounced like "sh"
If it was src/spec/ folks would likely think it was about specification, or ham, or something... 15:24
JimmyZ oh
jnthn Also it gives it a cute, short, name we can use to talk about it unambiguously. 15:25
JimmyZ I'm not native english speaker, so I'm more like optimize :P 15:27
15:39 raiph joined
raiph timotimo (or jnthn): Do you understand the basics of what's being said in the two page "1. Introduction" section, including the diagram, from the 1971 paper at cs.au.dk/~hosc/local/HOSC-12-4-pp381-391.pdf? If so, can you help me properly understand if/how it might one day be relevant to NQP and MoarVM? 16:07
(iiuc this stuff is a big part of the appeal of PyPy) 16:08
jnthn "One of the most difļ¬cult concepts in translator writing is the distinction between actions 16:09
done at translate time and those done at run time."
And don't I know it... :)
timotimo :3 16:10
jnthn BEGIN time in Perl 6 makes this especially fun :)
timotimo the fact that we can do so much in BEGIN time is pretty cool
do so much (without things blowing up left and right) 16:11
jnthn In the early days (pre-nom) you mostly just blew up there :) 16:15
Then we got QAST::WVals :)
Bah, e3c17bc was fail 16:17
timotimo how is that? 16:19
i'll probably see how in the commit you'll push to fix it :) 16:20
timotimo AFIrc 16:21
jnthn whoa, I think my deopt code mighta just worked... 16:25
dalek arVM/spesh: e1b09c0 | jnthn++ | src/core/op (2 files):
Fix deoptpoint instruction marks.
 16:36
arVM/spesh: 4942700 | jnthn++ | src/spesh/ (3 files):
Start computing de-opt mapping table.

Maps return addresses that may show up in specialized code to their original ones in the original code.
arVM/spesh: 496f417 | jnthn++ | src/ (4 files):
Store deopt mapping in cand, and cand in frame.
timotimo ah, the names didn't match 16:37
dalek arVM/spesh: eabbfd2 | jnthn++ | src/ (3 files):
Update Lexotic to be de-opt safe.

The effective handler set may change.		 17:32
arVM/spesh: 8b2f676 | jnthn++ | / (4 files):
Initial implementation of deopt.

Triggered by mixins. Fixes the NQP test suite, and gets Rakudo to survive the CORE.setting build with arg-type specialization.
jnthn Ohh. 17:36
It's not deopts fault
It's that istype analysis is broke
timotimo oops, my fault i guess 17:37
oh, right
i forgot to mention ... it should only ever do istype at specialize-time if the type to be checked against is "simple"
right?
though i'm not sure how to implement that in C-space
in nqp it'd probably be something like $type_to_check.HOW =:= ClassHOW
jnthn yeah, there's a type_check_with_cache which is clsoe but not quite it 17:39
Time for stroll/dinner 17:41
Will have a look at it afterwards
timotimo thank you 17:42
i thought type_check_with_cache wasn't a good choice as it would never say "no" 17:43
it would only say "probably not"
17:59 profdel joined
dalek arVM/spesh: f89bcb3 | jnthn++ | src/ (3 files):
Correct istype optimization.
 18:54
jnthn That almost gets it, but not quite 18:55
timotimo anything i can do about it? 19:01
jnthn oh... 19:03
I think it was setting the result wrongly.
oh, no...
It was right 19:04
Hmmm
timotimo MVM_TYPE_CHECK_NEEDS_ACCEPTS ā† i hadn't realized this exists; neat!
19:05 LLamaRider joined
jnthn oh, the issue may be something other than type check 19:05
yeah, it is
FROGGS dev/nqp$ ./nqp-m -e1 19:07
42
\o/ # I can haz embedded moarvm byteode thingies that get run im my executable! 19:08
jnthn :)
timotimo -e1? does that mean "run the first embedded bytecode" or something?
jnthn it's same as -e "1"
timotimo and the result of that is 42? :)
FROGGS *g* 19:09
jnthn uh... :)
FROGGS no, I compiled that "say(42)" to .moarvm, and included it as well as the nqp.moarvm
timotimo ah, ok :)
i like that 19:10
FROGGS++
FROGGS I need to disable that repl though, and tell it to just run the CUs
timotimo: I guessed :o)
when we link in the -lmoar statically, we would have a one-file thingy
timotimo i'll be interested to see how big that will turn out 19:11
FROGGS atm it is 2.2meg (nqp + say 42) 19:23
timotimo not bad 19:27
jnthn Turns out it's fine
It's just that a file earlier on in the build had been mis-compiled due to now-fixed bugs. 19:28
timotimo ah, what an annoying thing to track down!
the next thing i really ought to do in the spesh is to implement if_o and unless_o 19:39
they seem to be used much more frequently than if_i and unless_i
jnthn Well, but in contexts where we have the value? 19:40
timotimo only in that context
unless there are types where knowing the type of the thing would be enough to decide whether it's truthy or falsy? 19:41
jnthn No, you need the context 19:43
uh, the vlaue
timotimo good to know
jnthn And you can't do it in any case that does a method call
timotimo should not be terribly hard, then
there should be a simple flag i can read out to decide that, yes?
jnthn Well, the thing to look at is the boolification spec.
Oh, actually...there is quite a lot in that area.
Because the boolification spec *is* by type 19:44
And it tells you how to boolify it
so if you know the type you can remove a level of indirection potentially
timotimo oh, that's neat, too
jnthn I'm going to work my may now through things that lowered bindings do in Rakudo 19:45
And try to improve things a lot there.
timotimo \o/
wait, "work my may"? does that mean we'll have to wait another month? :( 19:46
jnthn *my way
haha...what a typo :)
dalek arVM/spesh: 7767e32 | jnthn++ | src/spesh/optimize.c:
Sort ops in order.

May improve code-gen.		 20:14
arVM/spesh: 7b11da1 | jnthn++ | src/6model/serialization.c:
Don't lose HLL owner during serialization.
arVM/spesh: b2dc797 | jnthn++ | src/spesh/optimize.c:
Optimize away hllize when possible.
timotimo how expensive is hllize in the normal case where the owner is already correct? 20:19
jnthn Not hugely, I think. 20:22
But it's still work we needn't do.
timotimo aye 20:23
20:55 flussence joined
dalek arVM/spesh: 0700217 | jnthn++ | src/spesh/ (4 files):
Specialize by type held within scalar container.

Enables us to start eliminating more type-check related instructions in Perl 6 binding.		 21:00
timotimo does nqp require some fiddling to make use of the same kind of optimization here? 21:02
jnthn It doesn't emit type checks; signature type checks only take effect for multi dispatch 21:03
timotimo ah, yes, now that you mention it i do remember that 21:05
dalek arVM/spesh: 30a287a | jnthn++ | src/spesh/optimize.c:
Optimize away isconcrete when possible.
 21:27
arVM/spesh: 629caee | jnthn++ | src/spesh/optimize.c:
Optimize away assertparamcheck when possible.
timotimo i'm quite fond of the fact that turning istypes and such into constant values will sometimes remove getting the value (for example the wval for the type itself) from the bytecode completely 21:34
jnthn Aww...somewhere along the line the Rakudo build has broken 21:41
timotimo should i try to build rakudo at a specific revision for you? 21:47
jnthn I...think it may have been a local Rakudo patch doing it :/ 22:12
timotimo i'll try to build moarvm/spesh then 22:13
jnthn Or partly to blame
timotimo are there juicy opportunities in the specializer right now to improve regex/rule performance? 22:14
/home/timo/perl6/rakudo/../install/bin/nqp-m --target=mbc --output=blib/Perl6/Actions.moarvm --encoding=utf8 \ 22:15
src/Perl6/Actions.nqp
Use of undeclared variable '$copy' at line 6366, near " {\n "
oh
that's a local patch, too :)
hmm 22:16
Stage parse : Package 'Scalar' already has parent 'Scalar'
at src/gen/m-Metamodel.nqp:663 (blib/Perl6/Metamodel.moarvm::43)
that's with all changes stashed
oh, huh 22:18
with MVM_SPESH_DISABLE=yesplease i get Stage parse : This type cannot unbox to a native integer 22:19
at src/gen/m-Metamodel.nqp:2673 (blib/Perl6/Metamodel.moarvm:compose:61)
jnthn b2dc797 seems to have been fine 22:22
Provided I make clean Rakudo first. 22:23
timotimo i make m-clean too
jnthn Hmm...0700217 also was after a clean 22:26
timotimo aye, b2dc works here, too
jnthn Trying 30a287a 22:27
aha...
timotimo Stage parse : Package 'Scalar' already has parent 'Scalar' 22:28
on 30a287a
jnthn Yup 22:29
I think that one must be to blame.
But...how on earth...
oh... 22:30
No, I don't see what's up 22:34
timotimo time to dig into the spesh logs :P
jnthn timotimo: Yeah. 22:36
timotimo it also breaks with MVM_SPESH_DISABLE
for you, too?
i couldn't get it to output any spesh log file either
jnthn Oh, I didn't try that, but that'd be a real weird thing
I mean, because if I get head
timotimo it is already weird to me.
jnthn And then comment out the optimize_isconcrete call 22:37
Then it's happy again.
timotimo i bet it's a missing break!
timotimo looks
nope.
jnthn No :P
dalek arVM/spesh: 02883cd | jnthn++ | src/spesh/optimize.c:
Back out optimize_isconcrete; it breaks something.
 22:38
jnthn I'm tired, but not *that* tired :)
jnthn pours a Russian Imperial Stout 22:42
spectest ain't bad with that 22:46
timotimo with the stout?
jnthn no, with 02883cd :P 22:48
timotimo oh 22:49
jnthn timotimo: If you fancy doing something to help spesh do better, getting the NQP regex optimizer able to play nice with the main language optimizer would be good
timotimo: Then we can remove the poisoning off all the regexes
timotimo: Which in turn means that spesh can compile-time resolve the methods 22:50
timotimo my $*OUTER_OPTIMIZER? :P
the action methods in particular, eh? 22:51
jnthn I think you alrady pass some adverbs to the regex optimizer. Maybe there is fine.
timotimo or does that also include calls to subrules?
jnthn No, it's already doing OK with the action methods
timotimo that's good
jnthn I mean the method calls to subrules, !cursor_start, etc.
timotimo that does sound helpful 22:52
why did i do the thing with @!outer 22:55
ah, a big part of that was moving stubbed out blocks somewhere so the compiler won't get sad during serialization/compilation 22:56
i wonder if it would be okay to directly call into $!outer_opt.visit_block 22:59
i wonder if i should visit the children of that block with the regex optimizer, too 23:00
23:01 colomon joined
jnthn timotimo: I think it should be safe 23:02
timotimo it doesn't explode at least 23:09
i'll have to spectest to convince myself it's all right
jnthn Did you unpoison the opt?
timotimo oh, i forgot that part :) 23:10
that would probably make a big difference
No contextual found with name '$/' - yeah, no :) 23:11
do i perhaps need to introduce the outermost QRegex to the union algorithm as something barrier-like? 23:15
without completely poisoning lowering?
jnthn Hm, not sure 23:18
timotimo this is kind of hard to debug :|
jnthn yes, I know, I did a bunch of it the other evening ;) 23:19
timotimo oh btw, is there something we can do if we know a dynamic var has been just declared and is now being used? can we get around the dynamic lookup in that case?
jnthn I think we already code-gen that smarter
As in, turn it into a lexical access
Enough hacking for today for me :) 23:20
'night
timotimo good night and thanks for all the fish^Wspesh!
23:30 btyler joined 23:35 colomon joined 23:54 colomon joined

Powered by Cro and the RakuĀ® Programming Language.

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

Thank you!