Indexing behavior on shape mismatches

[goretkin]

make sure to run this with more than just a single node.

assert(nprocs() > 1)
particles = dzeros(10*nprocs(),2);

println(particles[5,:])
println(particles[15,:])

#these should both work or not work.
particles[5,:]=[1,2]      #this works
particles[15,:]=[1,2]     #this doesn't work,
                            #error: in _F_: arrayref: index out of range

#these both work                                         
particles[5,:] = [1,2]' 
particles[15,:] = [1,2]'

[goretkin]

In my opinion, something like particles[5,:]=[1,2] should work even though size(particles[5,:]) != size([1,2]).

Also, particles[15,:] = 2 does work (it's equivalent to size(particles[15,:]) = [2,2]' )

[JeffBezanson]

There are more general issues here we haven't worked out.

How should using too few or too many indexes behave? Currently, if A is 2d, A[1,1,1] happens to work because the 3d ref method doesn't need to access size(A,3), but this is an accident. Should we pad sizes with ones, or require an exact match?

assign for Array currently accesses the right-hand-side array with linear indexing, so it is effectively doing a reshape. In fact, a vector can currently be reshaped to 2d using v[:,1:1], but you can't reshape it to 3d this way.

Generally our behavior is for the result of non-scalar ref to have rank equal to the number of indexes, but it doesn't always work. I'd propose we just allow this in full generality.