<sb0>
so it didn't install the one in conda.yaml this time, despite the constraint
<sb0>
(i was trying to build a 1.0rc2 again. it's a mess.)
<GitHub149>
[artiq] sbourdeauducq pushed 1 new commit to release-1: https://git.io/vVlHB
<GitHub149>
artiq/release-1 4d54695 Sebastien Bourdeauducq: conda: simplify llvmlite-artiq version requirement...
_rht has quit [Quit: Connection closed for inactivity]
<whitequark>
sb0: why did you *delete* the others?
<whitequark>
I have specifically left them there to be able to bisect any bugs across LLVM versions.
<sb0>
so that conda hopefully picks the right package on the buildserver without a build number specified
<whitequark>
could've moved them to a channel "deleted" or something, at least
<whitequark>
and now it's a mess. great.
<whitequark>
well, more mess, I guess.
<sb0>
I thought the LLVM upgrades were easy ...
<whitequark>
it has to be done and our infra isn't made for building old package versions
<whitequark>
it's all made for building just the latest one. building the old ones will require some kind of error-prone manual hack, like reverting it back and then reverting again.
<sb0>
also, i didn't think you were using conda for development
<whitequark>
the package that was installed is artiq-1.0rc1-py_82+git6a34a75
klickverbot has quit [Ping timeout: 246 seconds]
<whitequark>
which is actually not 1.0rc1, but the one built earlier from master
<whitequark>
this is incredibly dumb, because I install it as --use-local artiq. but that's broken too, apparently.
<whitequark>
fortunately, that's an easy fix, because I already fixed a stupid bug like that for the Windows unittests
<_florent_>
sb0: I'll look at the misoc sdram cleanup on Monday
klickverbot has joined #m-labs
<whitequark>
sb0: re C interpreter in HDL: yes, it is a terrible mess. I think there was some extremely simplified prototype?
<whitequark>
it didn't go anywhere, of course
<whitequark>
rjo: regarding the missing optimization. yes. not good. we can fork LLVM and implement the !dereferenceable_unconditionally attribute that would convey what we want.
<whitequark>
it's a one-line patch, if you don't count the boilerplate, anyway
klickverbot has quit [Ping timeout: 246 seconds]
<whitequark>
maybe it will be accepted upstream, but these attributes are incredibly messy already, so maybe also not.
<whitequark>
rjo: I can do it post 1.0 or pre 1.0, it will not take long to implement.
<whitequark>
and we should also mark the floating-point intrinsics as safe to speculatively execute, as I proposed before. which also requires temporarily forking llvm-or1k.
<rjo>
whitequark: forking llvm with those two changes is definitely fine by me. i guess that would remove all FP from the loop body?
<whitequark>
hm, let's try this...
<whitequark>
rjo: yes. that's the intent.
bb-m-labs has quit [Quit: buildmaster reconfigured: bot disconnecting]
<sb0>
wait, doesn't the buildbot also have the main channel
<sb0>
for dependencies?
<whitequark>
yes
<sb0>
why doesn't it install the package it just built, from the filesystem?
<whitequark>
obligatory: because conda is crap
<sb0>
what breaks exactly? automatic download of dependencies?
<whitequark>
hm
<whitequark>
can you show me the problem?
<whitequark>
which build is it?
<sb0>
maybe a stupid but functional solution is having a static list of dependencies hardcoded in the bot, then have conda install those, then install the package from the FS?
<whitequark>
let's not do that if possible at all. it's extremely fragile.
<sb0>
re-download sounds pretty fragile too
<whitequark>
there are enough points of failure that need manual care in this infra around conda anyway
<whitequark>
so, where is it redownloading things?
<sb0>
well I don't know, rjo ?
<whitequark>
he just went to the MTR
_rht has quit [Quit: Connection closed for inactivity]
<whitequark>
oh
<whitequark>
oh I know
<whitequark>
ah no, that was irrelevant
sandeepkr has joined #m-labs
evilspirit has joined #m-labs
rjo2 has joined #m-labs
<rjo2>
it's just that the highest Version in the Dev Label won't be from Master anymore.
<whitequark>
ah. yes. that's expected.
rjo2 has quit [Quit: AtomicIRC: The nuclear option.]