github.com/moarvm/moarvm | IRC logs at colabti.org/irclogger/irclogger_logs/moarvm
Set by AlexDaniel on 12 June 2018.
00:18 Merfont joined 00:19 Kaeipi left 01:18 Merfont left, Merfont joined 01:27 Merfont left, Merfont joined 04:14 Merfont left 04:15 Merfont joined 05:23 Util left 05:35 Util joined 07:00 bingos_ is now known as bingos 07:01 bingos left, bingos joined 07:03 bingos is now known as BinGOs 07:25 MasterDuke joined
nwc10 good *, #moarvm 07:32
07:52 brrt joined
brrt good * #moarvm 07:54
07:54 zakharyas joined
nwc10 good *, brrt 07:54
07:55 zakharyas1 joined
nwc10 brrt: I've been staring a lot at compiler assembler output. Is this somethign that will make me go blind? :-) 07:56
07:58 zakharyas left
brrt no, will not 08:04
mad, maybe
then again, might be a good lockdown time spender 08:05
nwc10 Yes, I think it makes me mad, maybe in two ways :-/
08:05 sena_kun joined
nwc10 at least, a bit disappointed 08:05
clang on arm manages to use LRDB to read one value
and then the other one, LDR
and then some ARM6 instrcution that I can't remmeber the mnemonic for to truncate *that* read down to 8 bits 08:06
dude, you just showed me that your code generator knows about LDRB - why are you now usign 2 instructions to do the other byte load when we both know you could use one?
gcc on ARM was not that stupid
probably not goign to change - I'm assuming that most clang ARM development comes from Apple, and Apple has gone 64 bit. 08:07
08:14 zakharyas joined
brrt probably yes 08:15
08:17 zakharyas1 left 08:39 sena_kun left 08:46 domidumont joined 09:31 sena_kun joined 10:13 brrt left
timotimo could be a memory alignment thing, so that the assembly fits more neatly on some boundary somewhere further down in the code 11:11
11:12 sena_kun left 11:17 zakharyas left
lizmat github.com/rakudo/rakudo/issues/39...-714389210 # posting here on Altai-man's behalf 11:22
12:25 zakharyas joined
Geth MoarVM/plain-char-can-be-unsigned: a989f7b553 | (Nicholas Clark)++ | src/strings/ascii.c
`char` can be unsigned. Rewrite the "is it ASCII?" test to handle this too.

Without this gcc correctly warns that the "comparison is always false" on platforms where plain char is unsigned, and the code (as was) was buggy.
On x86_64 (where plain char is signed) this commit makes no difference to the generated code.
13:01
MoarVM: nwc10++ created pull request #1363:
`char` can be unsigned. Rewrite the "is it ASCII?" test to handle thi…
13:02
14:18 sena_kun joined 15:14 domidumont left 15:34 patrickb joined 17:38 linkable6 left, evalable6 left 17:40 linkable6 joined, evalable6 joined 17:42 zakharyas left 18:39 rypervenche left
Geth MoarVM: a989f7b553 | (Nicholas Clark)++ | src/strings/ascii.c
`char` can be unsigned. Rewrite the "is it ASCII?" test to handle this too.

Without this gcc correctly warns that the "comparison is always false" on platforms where plain char is unsigned, and the code (as was) was buggy.
On x86_64 (where plain char is signed) this commit makes no difference to the generated code.
18:42
18:42 rypervenche joined 18:43 sena_kun left
samcv timotimo, MVM_string_buf32_can_fit_into_8bit ... MVMGrapheme32 val2 = ((active_blob[i] & 0xffffff80) + 0x80) & (0xffffff80-1); 19:21
so we do it 32 bits at a time. dunno how many bits we can do. 128? forget how wide simd are
timotimo there's even 256 wide ops 19:22
samcv well i mean it depends on the instruction but. i guess it could make sense for that i guess
though. actually i think it DOES vectorize it. i am remembering now
it has multiple alternatives based on how long the string is, and already does vectorize for wider vectors 19:23
assuming it still optimizes the same
timotimo it has the benefit that it doesn't have to check for \r\n and friends
samcv i remember it checks the length, does it x bits at a time, then if any remain does 32 bits for the rest 19:24
timotimo right, that makes sense
that's what our code is written for, or that's what gcc and clang do with it?
samcv well the code is written to do 32 bits at a time, though i wrote it with many tweaks until it actually did that properly. so i guess a bit of both 19:25
but verbatim it only says to do 32 bits at a time
timotimo OK 19:26
samcv MasterDuke, i don't think we get to memory speed limits unless we are just doing straight memcpy
19:30 brrt joined
MasterDuke not sure what you mean? 19:31
19:44 patrickb left 19:53 zakharyas joined 19:59 zakharyas left 20:19 brrt left
samcv sorry. referring to this comment "<MasterDuke> how many strings are we validating that don't fit in l2?" 20:30
also i'm curious about the FSA. what benchmarks show the greatest performance benefit from that? I am trying to compare using FSA to just using normal malloc, on linux and getting pretty much dead even. But guessing i'm not doing a good test case 20:31
MasterDuke well, i think our arenas might not be big enough. so the FSA falls back to malloc too frequently 20:35
but there was a recent change i made to concblockingques from malloc to fsa that made a noticeable difference 20:36
there is a rakudo issue with some commentary, don't remember the # 20:37
samcv: github.com/rakudo/rakudo/issues/3648 20:50
timotimo we can trace with systemtap or something how often what sizes happen in the FSA 21:02
MasterDuke i've logged it before, don't remember at all 21:04
most common value requested is 88, then 656, then 104, then 256, then 200, then 192, then 8 21:10
samcv: if you're interested, github.com/MoarVM/MoarVM/pull/1277 21:12
timotimo ideally we'd reduce the total amount of allocations 21:15
MasterDuke yeah, fewer allocations would be great 21:18
22:03 Kaeipi joined
MasterDuke most common bin is 10, then 81, the 31, then 12, then 24 22:06
22:08 Merfont left 22:11 Kaiepi joined 22:15 Kaeipi left
timotimo if we also count deallocations we can see how many of what size survive until the program exuts 22:22
MasterDuke most common bytes value for free is 88, then 656, then 256, then 104, then 200, then 192, then 8 22:35
timotimo macroify-allocs (or something) in the tools folder tries to give the allocations names on top of just sizes 22:38
it will make a very big diff that goes into most files 22:41
MasterDuke might get a chance to try tomorrow, off to sleep now 22:45