Use staged functions in @variable

[joehuchette]

Currently we don't recognize this as something we can store as a JuMPArray:

I = 1:3
@defVar(m, x[I])

It should be fairly straightforward to add a staged function in the generated macro code that takes the JuMPArray branch based on the type of I.

[mlubin]

As I alluded to in #337 we can't easily implement this until we move over to the dormant JuMPArray.jl implementation because we're currently generating custom types inside of the macros. That's sketchy enough as it is, I don't think we should try to generate types at runtime within staged functions.

[joehuchette]

Should we try to bump the ntuple inlining thing?

[mlubin]

Not sure if it will be helpful, it's already a well-known issue.

[mlubin]

Don't think this is needed for JuMP 0.10, it'll be easier to do after we drop support for 0.3.

[mlubin]

@joehuchette, are you planning on taking a stab at this soon? Does it need to block 0.14?

[joehuchette]

I meant to tag it 0.15. Thinking about this more, we won't be able to return a Vector{Variable} for my I=1:3 example above without checking the endpoint at runtime, which will ruin type stability. The best we could do with generated functions is to return a JuMPArray. Is it still worth it? If this is mostly about performance, rather than semantics, type instability seems like a show-stopper.

[mlubin]

We should be looking at the new arbitrary-indexed arrays for that: JuliaLang/julia#16260
So definitely after we drop support for 0.4

[chriscoey]

any new thoughts on this? #933 was added to 1.0 milestone, should this be also?

[joehuchette]

d032934