if anyone wants to crack into this https://clojure.atlassian.net/browse/ASYNC-247, would love a hand /cc @bronsa @hiredman

👀

vs would be overwritten as the loop loops, so it must be coll

this would be a phenomenal screen recording of how you debug it or perhaps an article if whoever fixes it is interested

I suspect https://clojure.atlassian.net/browse/CLJ-2145 may fix it already

yeah I can't repro with that fix applied

[~/src/clojure]> clj -J-Xmx32m -Sdeps '{:deps {org.clojure/core.async {:mvn/version "1.5.648"} org.clojure/clojure {:mvn/version "1.12.0-master-SNAPSHOT"}}}'
Clojure 1.12.0-master-SNAPSHOT
user=> (require '[clojure.core.async :as a])
nil
user=> (let [c (a/chan)] (a/onto-chan! c (range 1000000)) (while (a/<!! c)))
nil
user=>

(I can repro without it)

trying again in case it was just a lucky run..

yeah, I confirm the fix. I've just glanced at the core.async code so I can't offer a precise explanation, but the code shape looks "right" to hit the issue CLJ-2145 describes

I'm not saying you're wrong, but that doesn't match a couple things I saw in debugging

hmmm, though I didn't get an OOM in 1.12.0-alpha1 on this second run either, so hang on, maybe I keep getting lucky 🙂

so ignore this for now

the expansion of onto-chan's go-loop is something like this

Indexes 6/7/8 are the locals - you can see in lines 19-39 those closed over :once thunks are invoked, then saved and vs is saved as index 9 (and 10) later

in a debugger I see I think the original coll as the one that is retained (index 7)

yeah, I don't see it being nil'ed anywhere

so that's why I don't see a match to CLJ-2145 (no conditional, and different place where I see the head retained)

yeah the macroexpansion is clear about that

you both have more experience than me in here, I'm not sure what the analogous approach to locals clearing in jvm clojure would be - how easily can we know enough to not save those locals in the state array?

(separately, that expansion looks like it's redoing the resolution of all of the locals multiple times - ch, then ch and coll, then ch and coll and close?, a lot of this smells unnecessary)

I don't know how locals clearing works in the go macro

afaict, it doesn't :)

the enviroment re-creation stuff is meant to make it possible to emit big chunks of normal code that doesn't require rewriting as a state machine, there just aren't any big chunks here

there is a pass in tools.emitter.jvm to annotate locals at the point where they can be cleared. maybe I can try reusing that and emitting an aset to nil

tricky because it aliases things in locals instead of using direct refs to the state array

I am refreshing my memory on the ioc macro 🙂 so many subtleties I forgot about