| IRC logs at
Set by AlexDaniel on 12 June 2018.
01:11 ZzZombo left 01:16 lucasb left 04:18 coverable6 left, reportable6 left, quotable6 left, evalable6 left, benchable6 left, committable6 left, nativecallable6 left, squashable6 left, shareable6 left, bloatable6 left, statisfiable6 left, tellable6 left, greppable6 left, notable6 left, unicodable6 left, bisectable6 left, releasable6 left, nativecallable6 joined 04:19 coverable6 joined 04:20 greppable6 joined, statisfiable6 joined, shareable6 joined, bisectable6 joined 04:21 bloatable6 joined, quotable6 joined, committable6 joined 04:22 squashable6 joined, reportable6 joined, benchable6 joined, evalable6 joined, releasable6 joined, tellable6 joined, unicodable6 joined 04:23 notable6 joined 06:55 sena_kun joined 08:04 domidumont joined
nine For several hours, I've tried to come up with a reliable way to unsetup a native sub at all the required times and have failed so far. 08:58
But why do I need to do that in the first place? It's because MoarVM changes the state of the native call site without the upper layers noticing it. 08:59
But MoarVM already has a way to maintain and keep that state: serialization. It's just that this is not implemented for NativeCall reprs. Instead we rely on the upper layer to run initialization again. 09:00
But why is that? We actually have all the information we need to load that library and JIT compile the function body in the NativeCall's body. We just need to serialize that and run some initialization on deserialize. 09:03
MasterDuke sounds promising 10:15
nine And not just even works :) 10:26
At least for rakudo's tests and my golfed test case. There's still issues in Inline::Perl5, but that's just a matter of more debugging I guess 10:28
Oh, I commented out a line in Inline::Perl5 during debugging. That of course doesn't help at all... 10:30
There's one issue remaining that's really baffling. 12:30
Exactly one of Inline::Perl5's tests is failing with a "Stub code executed" error, i.e. one (and apparently really only one) of the native subs is losing it's CALL-ME magic without getting the optimized function body 12:32
13:02 Altai-man_ joined
nine Ok, how is it possible that the sub's invocation protocol gets changed but the code body not replaced when both is done by the !setup method? 13:02
The invocation protocol is a property of the type while the body is part of the object. Apparently only one of those gets repossessed (I guess the type) and thus changed by loading the precomped module. 13:04
13:04 sena_kun left 13:05 Altai-man_ left 13:13 domidumont left 13:50 ZzZombo joined 13:53 lucasb joined
nine I can circumvent this by not changing the invocation protocol and not replacing the code body during compilation. But that means that the function will never get the speedup as the setup code will only ever run once. 14:29
dogbert17 that sounds a bit bad doesn't it? 14:32
or perhaps I should have written suboptimal 14:34
nine Both are correct ;) 14:35
So....native subs get their special behaviour with the CALL-ME body mixed in. What if I mix in another role that gets rid of CALL-ME? 14:37
Instead of changing the Mixin after the fact
The question is: can a role hide a CALL-ME? 14:39
dogbert17 who might have the answer to that question? 14:42
nine As an alternative I thought, maybe it's enough to mixin any role. This way the routine would get a new type and I only change the invocation spec of this new type. This would mean that we could never end up with an object with a changed code body but with an unchanged invocation spec. 14:51
Alas....this only gives me a different repossession error: Incompatible MROs in P6opaque rebless for types Method+{NativeCall::Native[Method,Distribution::Resource]}+{Callable[NativeCall::Types::Pointer]}+{<anon|1>} and Method+{NativeCall::Native[Method,Distribution::Resource]}+{Callable[NativeCall::Types::Pointer]}
I think the whole issue is caused by repossession order. 15:15
dogbert17 so there are still bugs hiding :) 15:17
nine Not sure it can be called a bug. It may be a conceptual restriction.
I've been at this bug for 8 hours straight, so I'm not sure I've got enough brane left to explain this properly :/ 15:18
But....I wonder if there's another workaround worth trying: due to even weirder issues NativeCall has never created optimized function bodies during precompilation. But what if we could at least replace the stub code with a function body that runs !setup? I've had success doing this manually 15:19
dogbert17 is that 'easy' to do? 15:24
nine No, that'd require anything, literally anything to cooperate today... 15:27
Well....taking it one step further: if I already think about replacing the code body, why do we even add a CALL-ME in the first place? Why not replace the stub code with what CALL-ME is doing right now? 15:28
Then we'd never have to change the invocation spec in the first place 15:29
Feels like this whole thing is another exercise in finding out how deep a rabbit hole can really be 15:30
dogbert17 sounds clever. Are there any counterarguments, like code duplication?
nine No, it sounds desparate. Style is no longer an issue. This is about survival. 15:33
dogbert17 uh oh
timotimo i wish for programs to be runnable without the compiler getting loaded at all :S
at the moment we can't get that yet 15:34
hm. for cases like "empty code file, but a -M is provided" we could probably get away without loading Grammar and Actions 15:35
wouldn't that be nice
nine "An exception occurred while evaluating a constant" oh thank you! How helpful
dogbert17 sounds a bit like Visual Basic of old. "Array index out of range", and no we won't tell which array or where in the code that happened 15:36
nine: perhaps it's time for some food 15:47
nine Useless use of constant integer 1 in sink context (line 554) 15:52
No, actually this prevents "No such method 'sink' for invocant of type 'QAST::Stmts'"
Which would be triggered by literally any code after that if block. It's pure chance that this hasn't happened before 15:53
dogbert17: food does sound like a great idea 15:55
Why can't anything ever go smooth? 16:34
17:50 domidumont joined
timotimo - "find objects" shows BOOTHash objects grouped by size (it will of course later on also show what type is being looked for, lols) 17:55
19:08 domidumont left
nine m: perl6 -e 'class Foo { method foo() { ??? }; BEGIN { my $meth := Foo.^find_method("foo"); use nqp; nqp::bindattr($meth, Code, q<$!do>, nqp::getattr(-> |a { note "foo!" }, Code, q<$!do>)); }; };;' 19:37
camelia ===SORRY!=== Error while compiling <tmp>
Two terms in a row
at <tmp>:1
------> perl6 -e⏏ 'class Foo { method foo() { ??? }; BEGI
expecting any of:
infix stopper
nine m: class Foo { method foo() { ??? }; BEGIN { my $meth := Foo.^find_method("foo"); use nqp; nqp::bindattr($meth, Code, q<$!do>, nqp::getattr(-> |a { note "foo!" }, Code, q<$!do>)); }; };;
camelia Stub code executed
in method foo at <tmp> line 1
nine m: class Foo { method foo() { ??? }; do { my $meth := Foo.^find_method("foo"); use nqp; nqp::bindattr($meth, Code, q<$!do>, nqp::getattr(-> |a { note "foo!" }, Code, q<$!do>)); }; };;
camelia foo!
nine m: class Foo { method foo() { ??? }; BEGIN {; my $meth := Foo.^find_method("foo"); use nqp; nqp::bindattr($meth, Code, q<$!do>, nqp::getattr(-> |a { note "foo!" }, Code, q<$!do>)); }; };;
camelia Stub code executed
in method foo at <tmp> line 1
nine So.....WTH?!
Replacing the method's $!do at runtime is OK, but at BEGIN time it just doesn't do anything, except for when the method has been called at least once? 19:38
Ah, of course! It's just missing the magic invocation: 19:43
m: class Foo { method foo() { ??? }; BEGIN { my $meth := Foo.^find_method("foo"); use nqp; nqp::getattr($meth, Code, q<@!compstuff>)[1](); nqp::bindattr($meth, Code, q<$!do>, nqp::getattr(-> |a { note "foo!" }, Code, q<$!do>)); }; };;
camelia foo!
timotimo compstuff? %) 19:47
nine I've rarely written such self explanatory code
All tests successful. 20:26
timotimo nice
nine Without the CALL-ME there's no more need to fiddle around with the invocation protocol, i.e. there cannot be any divergence between the routine's type and its body. This should also remove the need to exclude clones (i.e. closures) from getting the speedups. 20:28
timotimo good 20:29
21:07 Ven`` joined 21:34 zakharyas joined 21:43 Kaiepi left 21:44 Kaiepi joined 21:55 zakharyas left 22:55 Ven`` left 23:00 committable6 left, notable6 left, greppable6 left, nativecallable6 left, shareable6 left, evalable6 left, bisectable6 left, squashable6 left, coverable6 left, statisfiable6 left, benchable6 left, releasable6 left, unicodable6 left, bloatable6 left, tellable6 left, quotable6 left, reportable6 left, evalable6 joined 23:01 releasable6 joined, squashable6 joined, unicodable6 joined 23:02 greppable6 joined 23:03 tellable6 joined 23:04 bloatable6 joined, quotable6 joined, shareable6 joined, reportable6 joined, statisfiable6 joined, benchable6 joined, notable6 joined, bisectable6 joined 23:05 coverable6 joined, committable6 joined, nativecallable6 joined
timotimo the collectable navigator now lets you use the browser's back and forward buttons 23:39