github.com/moarvm/moarvm | IRC logs at colabti.org/irclogger/irclogger_logs/moarvm
Set by AlexDaniel on 12 June 2018.
00:00 reportable6 left
Geth MoarVM/master: 5 commits pushed by (Timo Paulssen)++ 00:00
00:02 reportable6 joined, kiwi_65 joined
timotimo param_sn and param_sp are both :noinline, so now frames that use slurpy positionals can be inlined \o/ 00:14
and then of course there's the potential to scalar-replace small arrays and small hashes like the ones made by param_sp and param_sn respectively 00:16
just scalar-replacing hashes that never get assigned to or read from would already be a little win, so that they only get materialized when deopt happens 00:34
in many places, a slurpy hash parameter hash gets kept alive just by deopt points
Kaiepi what does spesh do? are there docs on it? 00:37
timotimo it logs usage information about all code that's being executed 00:39
then it finds code that's "hot" enough to be considered for dynamic specialization (that's where "spesh" comes from) 00:40
then it converts the bytecode of the "hot" frames into a form better suited for analysis (which is Static Single Assignment form, and splitting it into "basic blocks", linked together in a control-flow graph) 00:41
after that, information about how the code gets invoked (how many parameters, what types) and what was logged during execution is used to make optimizations to the bytecode. some are "certain", but many are "speculative" and require guards to be inserted into the specialized code 00:43
guards are cheap checks that allow the specialized code to jump back to unspecialized code to handle every general case
one of the most important optimizations that spesh does is speculatively inline other code
because it can figure out what happens to data that is passed from the caller to the callee and returned back out 00:44
and more recently, partial escape analysis came to moarvm, which allows to speculatively not allocate objects until they really really have to be an actual object. until then, their attributes and such just get mapped to local variables (locals map to registers on moarvm) 00:45
Kaiepi: does that help?
Kaiepi gimme a minute, i was afk 00:46
timotimo haha
jn's blog has loads of posts about spesh, and he's also given multiple talks that explain chunks of spesh
Kaiepi sounds like something i'd have to read to really understand 00:48
timotimo setting MVM_SPESH_LOG=/tmp/blah.txt and running some random code can give you a bit of an idea what happens
Kaiepi there's still a lot of moarvm i have yet to read. i've dealt with i/o, nativecall, and reprs fairly closely but that's about it 00:52
and parts of the jit i guess 00:53
the gc i'll need to pore through at some point before i finish my grant work so i don't screw up implementing async unix sockets 00:54
01:06 MasterDuke left 01:19 lucasb left 03:07 squashable6 left, squashable6 joined 03:10 kiwi_65 left 05:33 robertle left 05:55 brrt joined
brrt \o 05:55
nwc10 o/ 05:56
brrt timotimo: +1 on spesh-log-toolage 05:59
06:00 reportable6 left 06:01 reportable6 joined 06:19 fake_space_whale left 06:21 brrt left 06:41 rba left 06:42 rba joined 06:45 rba left, rba joined 07:05 domidumont joined 07:27 sena_kun joined 07:28 sena_kun left 08:16 robertle joined 08:40 zakharyas joined 08:45 sena_kun joined 09:22 zakharyas left 09:40 camelia left 09:42 camelia joined, zakharyas joined 09:55 brrt joined 10:01 zakharyas left 10:27 squashable6 left 10:32 brrt left, squashable6 joined 11:18 brrt joined 12:00 reportable6 left 12:01 reportable6 joined 12:14 sena_kun left, sena_kun joined
Geth MoarVM/master: 29 commits pushed by (Bart Wiegmans)++
review: github.com/MoarVM/MoarVM/compare/0...8283424598		 12:33
brrt ^ that was 'jit-expr-float'
timotimo cool 12:34
brrt bug reports at my address please... :-)
on to the next! 12:36
timotimo linear nodes, right?
12:51 domidumont left, domidumont1 joined
brrt well, either that, or devirtualization 12:54
basically, I aim to leverage the expression template precompiler, to minimze the amount of rewriting I'll have to do for the linear IR 12:55
jnthn
.oO( guards! )
 13:01
13:01 lucasb joined
brrt can't have them all at once 13:03
jnthn: ooc, do you mean, the deopt guards, or the spesh candidate guard tree? 13:05
different problems
jnthn deopt guards 13:07
Many times, those are what seems to terminate the expr JIT's fun
Though a few other things like sp_bind_o also seem to, which I guess is just missing templates
brrt was looking at sp_bind_o recently, but forgot what happened there 13:10
Ok.. I'll put it on my list... I actually don't think these should be hard?
and they might benefit from named templates 13:12
jnthn Well, for guards, in terms of code size what'd be *really* nice is if we just set in some defined register the deopt point and then branched to a label with some code at the end, or some such 13:18
So that we don't bulk out the generated machine code with lots of calls to the deopt function
And thus get better instruction cache usage
We can maybe even do a conditional write of the deopt index (or just always write it to a register unconditionally, if that's cheaper) 13:19
And then the code at an actual guard site becomes the guard + a (conditional) mov + a conditional branch, and if we don't deopt we don't take the branch.
brrt jnthn: 'call' sound suspiciously like the instruction you'd want :-D 13:22
for 'branch with a value in a predefined register'
and... I agree... and this is totally possible.
hmmmm. 13:23
not sure, btw, if we can do conditional moves on literal values 13:24
no, not with a literal
... but we can always just use 'rax' for this.... 13:33
since I have that as the spare register anyway
AlexDaniel timotimo: it did it again: colabti.org/irclogger/irclogger_lo...08-20#l136 14:13
timotimo: although your idea of not sending the messages if the user replies kinda works!
14:19 domidumont1 left 15:05 robertle left 15:21 brrt left 15:26 pamplemousse joined
pamplemousse o/ 15:27
15:40 sena_kun left 15:52 robertle joined 16:10 brrt joined 16:31 chloekek joined 16:39 brrt left 16:52 domidumont joined 17:15 brrt joined 17:39 brrt left 18:00 reportable6 left 18:04 reportable6 joined 18:39 MasterDuke joined 19:01 chloekek left 19:04 domidumont left 19:44 Ven`` joined
MasterDuke getcomp is tied for second place in causing number of `jit bailed completely` during a rakudo build. does it make any sense to try and jit it? or not, since it does the mutex locking? 20:12
20:13 pamplemousse left
MasterDuke it's tied with `param_on_o`. can the `param_on_*` ops be jitted? 20:14
20:14 Ven`` left 20:21 Ven`` joined 20:31 Ven`` left 20:39 chloekek joined
MasterDuke timotimo, et al.: any idea why valgrind complains that `Conditional jump or move depends on uninitialised value(s)` at github.com/MasterDuke17/libtommath...e067201R16 if you pass a malloced string instead of a calloced one? 21:15
21:59 robertle left 22:12 chloekek left
jnthn MasterDuke: I figur eyou know, but in case not: malloc doesn't zero memroy, calloc does 22:54
MasterDuke: Possible accident with failing to set the NULL terminator? 22:55
MasterDuke: Also, probably wise to use snprintf 22:58

Powered by Cro and the Raku® Programming Language.

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

Thank you!