brrt good * 15:06
i can't replicate the failure of MVM_box_str
on my machine :-(
MasterDuke have you tried a linux vm? you're on osx, correct? 15:10
would it help if i did a jit-bisect again? 15:12
brrt i'm on fedora 15:13
specially for this
:-( 15:18
you'll never hear me complain about bugs, but bugs i can't replicate are... madenning
*maddening
MasterDuke: if you would? 15:19
maybe only in combination with box_i
since i haven't enabled that one
i can replicate them when they're both enabled 15:28
(and it's bisectable. this is going to be weird) 15:32
brrt also, bugs in CORE.setting compilation suffer from being rather slow to bisect 15:39
MasterDuke bisect with both box_s and box_i? 15:48
afk for a while, but i started it running 15:52
timotimo we cache the envhash even though getenvhash is marked "usehll" 16:38
timotimo rather surprised to see that getenvhash is hardly encountered by spesh at all 17:38
perhaps the code that uses getenvhash should be changed to access it via nqp::getenvhash more often rather than binding to a variable 17:39
trying to get note("foobar") if nqp::getenvhash<LOUD> to disappear if LOUD is not set in the environment 18:01
there's so many left-overs %)
MasterDuke .tell brrt the box_s jit-bisect data is here gist.github.com/MasterDuke17/e74e1...256ba2487a 18:03
yoleaux MasterDuke: I'll pass your message to brrt.
MasterDuke timotimo: what do you mean left-overs? 18:06
timotimo bytecode that was involved in the check and call 18:18
MasterDuke ah 18:19
timotimo why do we have PHI nodes with only two arguments %) 18:20
YES! \o/ 18:22
all that's left from this whole ordeal is the const_s for the LOUD string
it's probably only left because of deopt boundaries 18:23
hm, the effect isn't quite as extreme as i'd hoped 18:26
7.7s before my patch, 6.9s after my patch
perl6 -e 'use nqp; for ^5_000_000 { say "test" if nqp::getenvhash<LOUD> };'
though with the debug output completely removed from the bytecode, i imagine it could press a frame or two below the inline limit 18:27
MasterDuke cool 18:29
timotimo we just need to rewrite every usage of debug switching in nqp and rakudo to use this pattern instead of assigning to a variable of some kind 18:30
because variable assignment could very well mess this up
(more tests surely needed) 18:34
unfortunately, my code breaks nqp's build rather early
MasterDuke timotimo: btw, have you ever looked at www.libtom.net/TomsFastMath/ ? 21:44
timotimo only briefly peeked at the main page 21:48
MasterDuke think there's any chance moarvm could use (at least some of) it? 21:50
timotimo probably 22:00
Geth MoarVM: Kaiepi++ created pull request #810:
Panic when Dyncall fails to allocate native callbacks		 23:27

Powered by Cro and the Raku® Programming Language.

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

Thank you!