samcv icc does seem to be faster 01:12
.tell brrt i'm getting lots of integer truncation and failing some integer tests in nqp using the ICC 01:23
yoleaux samcv: I'll pass your message to brrt.
samcv may or may not be related 01:24
these are the errors gist.github.com/samcv/daf1b850727a...64b10cf471 01:25
s/errors/warnings
not ok 1 - denormal 5e-324 is recognized and is not 0; not ok 7 - denormal 5e-324 is recognized and is not 0 (Uni) in t/nqp/115-nums.t 01:27
not ok 98 - super small result from div_In work in t/nqp/060-bigint.t
Geth MoarVM: 42d5d0220c | (Samantha McVey)++ | build/setup.pm
Add Intel C Compiler as an option to Configure.pl

This is a copy of gcc except with -wd858 which supresses a warning that const on a return type is meaningless (warning doesn't show up on other compilers).		 01:33
01:39 bisectable6 joined
japhb samcv: But is icc's warning *correct*? 01:45
Meaning, should we fix the codebase rather than silencing the warning?
geekosaur const on a return type means something different in gcc, iirc 01:48
actually, no. seems it's both meaningful and useful, but only relatively recently. I was thinking of __attribute__ (const) 01:50
samcv japhb: i'm not sure 01:51
geekosaur stackoverflow.com/questions/871633...nst-value# (links to a duplicate which doesn't include all of the discussion) 01:52
samcv though i'm more concerned with the warning in the gist gist.github.com/samcv/daf1b850727a...64b10cf471
and that it fails some tests :\ though that could be unrelated
geekosaur warning seems correct on 64-bit, unless you're doing the mixed-model thing a pointer is 64 bits, not 32 01:53
oh, sorry, code actually expects that 01:54
so warning is correct but truncation is wanted there
samcv i'm guessing maybe something Zoffix did make it not portable or something 01:56
gonna run a full rakudo spectest now
geekosaur no, it's portable, and it's being done appropriately and explicitly
samcv no i mean other code 01:57
geekosaur well, technically you;d mask off the high bits I guess just to be completely clear, but that's a bit wasteful
samcv not the one i linked about the warning, but the changes zoffix made to fix denormals, i'm guessing it may not work for all compilers. i'm not sure
yeah
geekosaur the denormals may be a different issue I was going to comment on next
samcv ok :)
geekosaur gcc etc. these days assume you want full internal precision from floating point
icc may not 01:58
there;s some gcc options that will disable full precision, which may help with diagnosis; I'm trying to dig them up 01:59
samcv looking at the manpage. there's Set internal FPU precision to 24 bit significand, and 53 bit or 64 bit 02:00
^ talking about ICC
geekosaur huh. maybe icc also uses full precision by default 02:03
and in fact it looks like gc may still not do so; I thought that had changed, but sse2 and later more or less force it anyway. -ffast-math would get you 870 bit precision in that case
samcv hmm so it passes spectest 02:04
geekosaur which you might try with gcc to see if it then has the same or similar test failures. (might also need to disable -msse2 or whatever)
er, 80 bit precision
02:04 Merfont joined
geekosaur some discussion about floating point stackoverflow.com/questions/729586...ode-in-gcc 02:06
samcv also it could be due to it using icc's math libraries
which i think it uses
i'm gonna try the -pc80 flag to icc and see if that changes anything 02:07
nope. still the same 02:08
5e-324 is 0 though so maybe something is weird 02:09
geekosaur that's using higher precision than gcc does 02:10
samcv well it works fine on clang and gcc with MoarVM's setup 02:11
geekosaur but it looks like higher precision in the exponent, not the mantissa
samcv hm
it could be in strtod maybe? 02:13
geekosaur conceivable but I'd say it's a fairly severe bug because 5e-324 shouldn't even be possible with (double) 02:14
samcv so the test is maybe wrong?
geekosaur at least I think it's not; not sure where float limits are in gcc
samcv geekosaur: ok it does seem to be strtod that is returning a different result 03:33
04:18 evalable6 joined
samcv 5e-308 works but higher exponents don't 04:30
04:34 Zoffix joined
Zoffix 5e-324 is possible. It's a denormal, where portion of mantissa is used for the exponent 04:34
en.wikipedia.org/wiki/Denormal_number
geekosaur oh, that would break strtod, yes 04:35
if it's not expecting that
Zoffix "<samcv> you are likely interested in problems with denormal support on the intel c compiler".... 04:36
Not really. 04:37
04:37 Zoffix left
samcv yeah i think that may be correct 04:39
ok it sets errno to ERANGE and i guess we don't check errno 04:49
05:24 domidumont joined
samcv ok yay i fixed it 05:27
05:30 domidumont left
Geth MoarVM: 2052888eba | (Samantha McVey)++ | build/setup.pm
Fix handling of denormals for Intel C Compiler

Without setting the `-fp-model precise -fp-model source` CFLAGS code isn't compiled to handle denormal floating point values. This caused a few nqp tests for denormals to fail.		 05:30
05:30 domidumont joined
samcv geekosaur: thanks for the hints and helping me get this solved 05:31
06:22 domidumont left 06:31 lizmat joined 06:46 lizmat left 06:48 domidumont joined 06:59 lizmat joined, domidumont left 07:35 zakharyas joined 07:36 zakharyas left 07:37 zakharyas joined 07:56 zakharyas left, zakharyas joined 08:00 zakharyas1 joined 08:01 zakharyas left 08:15 zakharyas1 left 08:21 zakharyas joined 08:26 zakharyas left 08:32 zakharyas joined 09:11 zakharyas left 09:37 Redfoxmoon joined 09:48 zakharyas joined 10:44 brrt joined
brrt \o 10:44
nwc10 o/ 10:45
10:47 robertle joined
brrt what's up with the integer truncation business 10:54
lizmat looks like Rakudo GH #1905 is more of a moarvm issue 11:24
11:26 dogbert2 joined 11:46 lizmat left
brrt .tell samcv that's normal, that's just what dynasm does 12:11
yoleaux brrt: I'll pass your message to samcv.
13:23 ggoebel joined 14:05 zakharyas left, zakharyas joined 14:09 zakharyas left 14:57 domidumont joined 15:02 brrt left 15:29 brrt joined 15:35 domidumont left 15:54 ggoebel left 16:00 robertle left 16:08 lizmat joined 16:09 Redfoxmoon left, Redfoxmoon joined 16:25 ggoebel joined 17:01 brrt left 17:02 brrt joined
samcv ok cool that's what geekosaur i think said too 17:03
yoleaux 12:11Z <brrt> samcv: that's normal, that's just what dynasm does
brrt :-) 17:08
17:18 brrt left 18:07 ggoebel left 18:23 ggoebel joined 19:04 lizmat left, lizmat joined 19:12 brrt joined 19:23 lizmat left, lizmat joined
brrt i'm going to humbly request attention to my PR: github.com/MoarVM/MoarVM/pull/868 19:40
timotimo i've only read the description, but i say :+1: :P 20:17
brrt :-) 20:19
timotimo what's with the whitespace change in Configure.pl's line 371 20:20
looks like the : is supposed to go in the same column as the ? above it
... and the : on the line before that is also too far to the left?
brrt i dunno, that may have been cperls' work 20:21
20:26 zakharyas joined 20:29 lizmat left 20:37 brrt left
samcv .tell brrt it seems slower by the testing i've done 21:24
yoleaux samcv: I'll pass your message to brrt.
21:25 zakharyas left
samcv 4.54 before after 4.60 for this thing 21:32
dd if=/dev/urandom bs=100 count=800000 | base64 > /tmp/stuff; perl6 -e 'use nqp; my $hash := nqp::hash(); my @file = "/tmp/stuff".IO.slurp.lines; my $t1 = now; for @file { nqp::bindkey($hash, $_, 1) };my $t2 = now; for @file { nqp::atkey($hash, $_) }; '
MasterDuke we should get that moarvm benchmark page up and running again 22:45
...which i can never remember the address of 23:07
huggable: bench 23:23
huggable MasterDuke, github.com/japhb/perl6-bench
MasterDuke japhb: ping 23:41
japhb Pong 23:46
You wanted to know something about perl6-bench? 23:47
I wrote the code there; for a while jnthn was running it daily on one of his servers and posting the results, but I've not done that.
Note that many perl6-bench tests -- especially the microbenchmarks -- are probably wildly out of date WRT to the current state of Rakudo optimization (or lack thereof, as the case may be). 23:49
MasterDuke so far i've had to add a .cache on line ~75 of bench 23:54
but things seem to be running otherwise
i also removed most entries from components.json, but that's just because i only wanted to run it for moarvm and perl5 23:55
hm, bunch of Configure.pl errors, might be i need some perl5 module installed though... 23:58
japhb Yeah, that I would believe.
MasterDuke build::setup isn't a core module? 23:59
japhb BTW, you don't *have* to remove things from components, the driver script is relatively forgiving in different ways to indicate you only want to work with a subset of available engines/builds.

Powered by Cro and the Raku® Programming Language.

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

Thank you!