Julia: strange error for deprecated scalar-fill indexed assignment

[JeffBezanson]

With the deprecation removed, attempting a[idxs] = scalar gives an unclear error:

julia> a = fill(1, 10);

julia> a[1:5] = 0
ERROR: MethodError: no method matching setindex_shape_check(::Int64, ::Int64)
Closest candidates are:
  setindex_shape_check(::AbstractArray{#s57,1} where #s57, ::Integer) at indices.jl:218
  setindex_shape_check(::AbstractArray{#s57,1} where #s57, ::Integer, ::Integer) at indices.jl:221
  setindex_shape_check(::AbstractArray{#s57,2} where #s57, ::Integer, ::Integer) at indices.jl:225
  ...
Stacktrace:
 [1] macro expansion at ./multidimensional.jl:641 [inlined]
 [2] _unsafe_setindex!(::IndexLinear, ::Array{Int64,1}, ::Int64, ::UnitRange{Int64}) at ./multidimensional.jl:636
 [3] _setindex! at ./multidimensional.jl:631 [inlined]
 [4] setindex!(::Array{Int64,1}, ::Int64, ::UnitRange{Int64}) at ./abstractarray.jl:998
 [5] top-level scope at none:0

[acrosby]

It also seems that the .= approach, while intuitive if you know about it, is not particularly well documented. Should probably be included in the indexing and assignment section of the arrays page?

[BoundaryValueProblems]

I'm wondering what is the status of this problem. We also encountered this problem. Definitely, it is a good idea to include more explanation about indexing ans assignment. Many people want to do something like:

julia> x=randn(10);
julia> x[x .> 0.0] = 0.0
ERROR: MethodError: no method matching setindex_shape_check(::Float64, ::Int64)
Closest candidates are:
  setindex_shape_check(::AbstractArray{#s57,1} where #s57, ::Integer) at indices.jl:218
  setindex_shape_check(::AbstractArray{#s57,1} where #s57, ::Integer, ::Integer) at indices.jl:221
  setindex_shape_check(::AbstractArray{#s57,2} where #s57, ::Integer, ::Integer) at indices.jl:225
  ...
Stacktrace:
 [1] macro expansion at ./multidimensional.jl:641 [inlined]
 [2] _unsafe_setindex!(::IndexLinear, ::Array{Float64,1}, ::Float64, ::Base.LogicalIndex{Int64,BitArray{1}}) at ./multidimensional.jl:636
 [3] _setindex! at ./multidimensional.jl:631 [inlined]
 [4] setindex!(::Array{Float64,1}, ::Float64, ::BitArray{1}) at ./abstractarray.jl:998
 [5] top-level scope at none:0

Of course, if I replace the scalar 0.0 by a proper size array, e.g., zeros(sum(x .> 0.0)), i.e., then it works, but it's cumbersome really:

julia> x[x .> 0.0] = zeros(sum(x .> 0.0))
4-element Array{Float64,1}:
 0.0
 0.0
 0.0
 0.0
julia> x'
x'
1×10 Adjoint{Float64,Array{Float64,1}}:
 0.0  0.0  0.0  -0.565677  -1.90141  -0.441064  0.0  -0.305993  -0.178488  -1.72595

It would be really nice if this scalar assignment works as before in Julia v0.6.x and 0.7.

[fredrikekre]

You are supposed to use x[x .> 0.0] .= 0.0.

[BoundaryValueProblems]

@fredrikekre ; thank you very much for your clarification!

[BioTurboNick]

I agree that a better error message would be helpful here. I imagine this will be a common novice mistake. Suggesting broadcast operators would give the user a cue.

[dourouc05]

I've just hit this problem while migrating a part of my code from Julia 0.6 to 1.0 (for now, 1.0.1), and it took me a few hours to track it down. A better error message would be much welcome!

I would be willing to propose a PR for this, but I have no idea how you could implement it…

[mbauman]

That'd be great @dourouc05! I _think_ it might be a simple as defining setindex_shape_check(::Any...) = throw(ArgumentError("better message"))). You can try it interactively with:

julia> @eval Base setindex_shape_check(::Any...) = throw(ArgumentError("nice string"))
setindex_shape_check (generic function with 7 methods)

julia> a = fill(1, 10);

julia> a[1:5] = 0
ERROR: ArgumentError: better message

Make the message better, test more cases, add those cases as tests, and you'll be on your way!

[dourouc05]

@mbauman : I just did this with #31085. What would be the next steps?