WIP/RFC: Converted WeakKeyDict to hash & compare with object-id

[mauro3]

This implements Stefan's suggestion to use object-id hash in the WeakKeyDict: #24941 (comment), which has currently a "Triage" label. Maybe this PR can help inform that decision. The changes were pretty minimal and no breakage occurred.

TODO if we go ahead with this:

• run package evaluator(?)
• rename WeakKeyDict to WeakKeyIdDict
• deprecate
• NEWS item

I can do this on Friday if triage is positive.

TODO if we don't go ahead with this:

• add a test for key-conversion and the key-dropping (WeakKeyDict should not convert keys as otherwise it can drop them #24941)

x-refs: #3002, #24941, #28182

[mauro3]

There was one relevant CI failure on Travis Ubuntu 32bit: https://travis-ci.org/JuliaLang/julia/jobs/405256513, essentially:

    A = [1]
    AA = copy(A)
    wkd = WeakKeyDict(A=>1, AA=>2)
    @test length(wkd)==2 # passed
    kk = collect(keys(wkd))
    @test kk[1]==kk[2] # errors with BoundsError: attempt to access 1-element Array{Array{Int32,1},1} at index [2]

This is a bit odd as length(wkd)==2 but length(collect(keys(wkd)))==1. Any ideas? I don't have a 32bit system at hand, so it will be a bit tricky to debug. Also, the other 32bit builds were fine.

[vtjnash]

GC.@preserve A AA begin
...
end

[mauro3]

Thanks, fixed! Quesion: why would they need to be preserved? The do not go out of scope and thus should not be GC'ed, no?

[StefanKarpinski]

A is used subsequently so is relatively safe but AA isn't so it could be garbage collected.

[mauro3]

The == is actually needed: #24941 (comment).