I'm not certain this is correct, but here's my understanding: `a =
ImagePos(img, 10, 20)` absolutely must protect `img` from being GCed in code
like this:
function foo()
img = rand(5,5)
return ImagePos(img, 10, 20)
end
But currently objects on the heap do not seem to protect stack-allocated
objects from being GCed. So, the current choice is stack-allocate all(?)
immutables that are non-bitstypes, just to make sure we adequately protect
objects from being GCed. IICU, the new PR adds the ability to support
protection from heap-allocated objects, which means that it will no longer be
necessary to stack-allocate non-bitstypes immutables.
Best,
--Tim
On Sunday, February 07, 2016 06:39:51 AM Cedric St-Jean wrote:
> There's an issue <https://github.com/JuliaLang/julia/pull/12205> on Github
> to add GC support for stack-allocated objects, and that makes no sense to
> me! Could someone please help me out? In my mind, stack-allocated objects =
> Int+Float+...+Immutable (in some circumstances). I thought that with
> Immutables, if I have
>
> immutable ImagePos
> img::Image
> x::Int
> y::Int
> end
>
> function blah(img)
> a = ImagePos(img, 10, 20)
> foo(a)
> l = Any[40, a]
> end
>
> then `a` is stack-allocated. The foo(a) call may either copy `a` further,
> or just pass on a stack pointer to the existing `a`, depending on compiler
> details. Any stack-allocated object gets automatically wiped as the stack
> unwinds, hence does not need GC'ing.
>
> Then on the `l` line, because it's an Any array, the ImagePos needs to be
> copied, boxed and heap-allocated. So that one needs GC.
>
> What did I miss?
>
> Cédric
