jl_array_del_beg and friends refuse to resize an array when only temporary data is actually shared

[carlobaldassi]

Example:

julia> splice!(vec([x+y for x in [1,2], y in [3,4]]), 1)
ERROR: cannot resize array with shared data
 in splice! at array.jl:771
 in anonymous at no file

Since the matrix is temporary, no data is actually shared, and the call should succeed.

Related issue: #4311

[JeffBezanson]

Unfortunately, while being "shared" is part of the array abstraction, being "temporary" is not.

[JeffBezanson]

Actually, we can probably handle the case described here, where the reshaped array is 1d. If the original array is 1d, we can't grow it if other arrays use its data. But in this case we could just make a new copy.

[carlobaldassi]

I still think this should be made to work somehow though. Breaking up things to make them more explicit:

julia> a = vec([x+y for x in [1,2], y in [3,4]])
4-element Array{Int64,1}:
 4
 5
 5
 6

julia> gc()

julia> splice!(a, 1)
ERROR: cannot resize array with shared data
 in splice! at array.jl:771
 in splice! at array.jl:760

Here, when splice! is called, a surely is the only owner of the data, so an error claiming that there is shared data is surprising, at least.
In other words, if being "shared" is part of the abstraction, I'd argue that it should keep track of when shareholders have died out; this doesn't actually involve the notion of "temporary" array, at least not directly.

[StefanKarpinski]

This starts to become reference counting if we go much beyond what we're doing now.

[JeffBezanson]

We could keep a reference count of sharing arrays, but it seems strange to me to have a splice! work after a gc but not before. Then the behavior would appear effectively random.

[StefanKarpinski]

Yeah, that's pretty wonky.

[carlobaldassi]

Sorry, I didn't mean that it should work only after gc. I surely don't know anything about how the internals work, but my reasoning was: if gc knows that an array is no longer alive, and it's the one preventing splice! from working, it should be possible to use that information. I don't know if this ends up doing reference counting.

[JeffBezanson]

The way the gc determines that information is by performing a garbage collection. There is no other way to find out.

[carlobaldassi]

I see. Is ref count of sharing arrays feasible, or should we close this as wontfix?