Missing GC.@preserves?

[tkoolen]

Cbc.jl doesn't have a single GC.@preserve statement, but post-0.6, I really think it should. What's preventing prob from being garbage collected in calls like

https://github.com/JuliaOpt/Cbc.jl/blob/220f0fa5d94a4382261e76fc473ec494dd85382a/src/CbcCInterface.jl#L270

?

[mlubin]

Last time I looked into this, I understood that function arguments are safe from GC collection because they can't be freed in the calling scope until after the function returns. Has this changed?

[tkoolen]

Even if that's the case (I personally haven't read that anywhere) prob.p (i.e., getproperty(prob, :p)) is the argument, prob itself isn't a ccall argument and may not be used anywhere else, so its finalizer,

https://github.com/JuliaOpt/Cbc.jl/blob/master/src/CbcCInterface.jl#L77

may be called.

[tkoolen]

Sorry, I didn't fully parse your message.

they can't be freed in the calling scope until after the function returns

This is definitely not true anymore in >0.6.

[tkoolen]

See e.g. https://discourse.julialang.org/t/properly-using-finalizer-ccall-cconvert-and-unsafe-convert/6183/10?u=tkoolen.

[tkoolen]

Maybe you read https://discourse.julialang.org/t/julia-v0-6-vs-v0-7-gc-behaviour/10002/6, but if that's still accurate, it's precedented on @noinline.

Edit: and see also https://discourse.julialang.org/t/julia-v0-6-vs-v0-7-gc-behaviour/10002/8.

[mlubin]

Ok, I understand why function arguments wouldn't be special if the function is inlined. Thanks for the links.

[rschwarz]

I have not read all the links completely, but are you worried that prob will be GC'ed during, after, or even because of the call to solve, even though prob is still referenced in some outer scope?

[tkoolen]

A problem is that prob::CbcModel might actually not be referenced anymore in some outer scope after one of the Cbc function calls, at least not in any GC-visible way, which as of right now allows Julia to potentially run the CbcModel finalizer before the ccall, and the ccalls expect prob.p to be a valid pointer.

[rschwarz]

Thanks for the clarification. I still can not imagine a scenario where this reference to prob is missing or "invisible", but I will read through the links in detail now.