deepcopy duplicates arrays

[denizyuret]

a=rand(3,5)
b=(a,a)
b[1]===b[2] => true
c=deepcopy(b)
c[1]===c[2] => false

To fix this issue, on line 59 of deepcopy.jl we can replace:

        return copy(x)

with

        return (stackdict[x]=copy(x))

[StefanKarpinski]

This does seem to be a consistent change with the premise that deepcopy should be equivalent to serialize and deserialize.

[JeffBezanson]

Agreed, though currently the serializer does not accurately record shared references to bits arrays. I found it impossible to get acceptable performance with strings when doing that.

[StefanKarpinski]

Could we special case strings since they're mutable in implementation but immutable in usage?

[timholy]

I use deepcopy in Images precisely because it copies arrays: JuliaImages/Images.jl#356 (comment). I guess I've always thought that deepcopy's purpose is to give me a totally-independent version of something, so that if I mutate it the original will be unchanged.

Should Dicts instead make copies of all their keys and values? Or should I write a custom copy function?

[StefanKarpinski]

Would your use-case really be messed up by not copying strings? That seems pretty benign to me since one shouldn't mutate the storage of strings.

[timholy]

It's more the "value" than the "key" I'm worried about. E.g., "pixelspacing" => [2mm, 2mm] you don't want to change the original if you set img["pixelspacing"][2] = 3mm.

[denizyuret]

I am a bit confused: I was not proposing any sharing between the original
and its deepcopy. It is just that if two things are shared in the source,
they should be independently created but in a shared way in the deepcopy.
For example if a=rand(3,4) and b=(a,a), when we do c=deepcopy(b), there is
no sharing between b and c, just between c[1] and c[2].

best,
deniz

On Sat, Sep 26, 2015 at 4:09 PM Tim Holy [email protected] wrote:

It's more the "value" than the "key" I'm worried about. E.g., "pixelspacing"
=> [2mm, 2mm] you don't want to change the original if you set img["pixelspacing"][2]
= 3mm.

—
Reply to this email directly or view it on GitHub
#13124 (comment).

[timholy]

I didn't read the test case carefully enough. Yes, that's a problem. Carry on 😄.