subbu has quit [Quit: Leaving]
drbobbeaty has quit [Ping timeout: 256 seconds]
drbobbeaty has joined #jruby
drbobbeaty has quit [Ping timeout: 250 seconds]
drbobbeaty has joined #jruby
MattPattersonGit has quit [Quit: You have been kicked for being idle]
<headius> ok more good news, FFI specs are green except for the two that pass a Ruby Function object to C as a callback
<headius> I need to see if that passes on CRuby FFI because it could be tricky with M1 security requirements
<headius> (can't generated an executable endpoint at runtime without jumping through some hoops)
<headius> enebo: in between running M1 tests and cleaning out the hamster cage I am looking at
<headius> the cuplrit is the RECache in BuildDynRegExpInstr... it is caching even when it isn't a "once" dregex
<headius> this code goes back to 2015 so this bug has been out there for a long time
<enebo[m]> ah I marked for 9.3 but I guess it will be fixed on both
<headius> I don't think it can cache anything if it is not /o
<enebo[m]> I think this also changes for 3.0
<enebo[m]> but this is obviously broken for 2.5/2.6
<enebo[m]> The cache could see if it is the same source string and keep caching it perhaps?
subbu has joined #jruby
<headius> enebo: I don't see much value in that
<headius> typically a dregexp should not be the same string repeatedly so it wouldn't help many cases, and we'd need to cache the strings and do a comparison every time
<headius> the simple fix is to align with jit and just skip cache for non-once regexps
<headius> ok I guess we are way ahead of the curve now... the FFI gem does not build properly for darwin-arm64 and I'm not sure how to fix it
<headius> I don't grok extconf
<headius> so they may or may not pass this spec but as near as I can tell it has never been tested on M1 properly
<headius> enebo: I'm inclined to skip these two specs on M1 since it will take some research to figure out how to support C function callbacks from Java/Ruby
<headius> I don't believe any Ruby functionality we ship depends on the feature
<enebo[m]> sure it is several steps forwards in any case
<headius> looks like some finesse required to disable write protections to executable memory when generating the stubs, and god knows how that will work across threads
<headius> so yeah, I think this is ok to punt for now and maybe add some logic to FFI::Function failing fast if you try to use it on darwin-aarch64
<enebo[m]> MRI does it for their FFI and JIT right?
<headius> I doubt it
<enebo[m]> not for mjit and yjit does not run on M1 so not the jits
<headius> I don't think fiddle supports callbacks and they probably aren't testing the jits on M1 yet
<enebo[m]> mjit loads a dll
<enebo[m]> oh ok
<headius> and the FFI gem doesn't even build right for M1
<headius> we're on our own
<enebo[m]> well if they don't have it yet then maybe it is not a big issue
<headius> GraalVM doesn't support this platform yet either
<headius> JRuby blazing the trail again
<enebo[m]> Adding that rocket felt funny inside
<headius> hahah
<headius> enebo: when you get a change try to run spec:ffi for me
<headius> I'm getting weird errors locally
<enebo[m]> jruby-9.3 ran it fine
<headius> ok something must be messed up in my local copy, rubygems complains about some file being "owned by" my OS and not by rubygems
<headius> the fs mode and owner looks fine
<headius> it's working fine on M1 so 🤷‍♂️
<headius> mvn -Ptest is green on M1 CI
<headius> spec:ffi is green as well, minus those two skipped specs
<headius> weird noise out of the spec:ruby:fast job but it hasn't failed yet
<headius> if that passes we will have three key runs green on M1 and I will merge the branch into 9.3
<headius> this should be green but the closest it got had some really weird failure in a Hash#hash spec
<headius> something is weird about the spec:ruby:fast run via CI... it shows way more failures than running in a terminal and doesn't complete
<headius> there's definitely some networking oddities on MacOS instances because now a UDP spec is hanging there
<headius> could be some multicast that isn't supported or just isn't working with this cloud network
<enebo[m]> yeah I could see a network limit multicast
<headius> I'm going to tag it off like the hanging unix socket specs but we'll want to revisit these
<headius> I did get a green spec run on my personal M1 machine so this seems specific to the macstadium env
<headius> enebo: same question as last week... what's our plan for 9.2 vs 9.3 branches?
<headius> this dregexp thing affects both
<headius> normally I'd do a PR to 9.2 and merge down but 9.2 has no CI currently
<enebo[m]> 9.2 can just be cp'd back from 9.3
<enebo[m]> we decided to not merge forward on 9.2 any more right?
<enebo[m]> but we will to master from 9.3
<headius> that's fine with me, I wasn't sure we'd decide
<headius> decided
<headius> fixes to 9.2 have reduced to a trickle at this point
<enebo[m]> I think we did. I think the conclusion is 9.2 is low risk and its own thing now
<headius> ok I got a completed spec run on this M1 CI instance with four other network fails (looks like port permissions etc)
<headius> so it does appear to be greenish when run directly, modulo env weirdness
<headius> oh this local issue running specs is due to my regexp patch, I guess it's not right
<headius> enebo: PRs pushed for both branches for
<headius> so it wasn't actually caching the dregexp for non-"once" but it was using the same shared field as a holder
<headius> when the dregexp is not "once" it could then be changed across threads, which caused the bug
<headius> I am very surprised nobody reported this before
<headius> or maybe they did and could not come up with a small repro
<headius> brb, lunch-ish
<headius> ok, I have a bit more time today to look at the other targeted issues
<headius> enebo: what aren't you looking into?
subbu has quit [Quit: Leaving]
<headius> that is causing some jobs to fail on our end, not our fault
<headius> the ones that ran cover all suites and seem fine so I'm merging
<headius> enebo: hah because Travis gave us some juice the 9.2 branch has CI again:
<headius> we should just punt on moving 9.2 to GHA since this is working again
<headius> enebo: I punted the remaining 2.6 thing to 9.3.4 and detargeted that date week parsing draft PR that has been idle since 2020
<headius> also merged in the two working jobs on M1 but I will keep the branch around for ongoing improvements
subbu has joined #jruby