Fix Issue #5985: sprandbool() fails and does not have unit tests

[jperla]

Fixes an issue where sprandbool fails

[andreasnoack]

The problem is that for efficiency reasons the fields should be concrete types not abstract. I think that it is better that nzval is a Vector{Bool} instead of a BitArray{1} for sprandbool.

[jperla]

But then that will be less efficient for boolean matrices.

Shouldn't the JIT take care of this?

[jperla]

@andreasnoackjensen what if I add a new Union type of BitArray{1} and Vector{T}?

[tkelman]

A boolean sparse matrix doesn't really need a values array at all, the nonzeros are implicitly true. Can you express whatever sparse boolean matrix operation you're looking to do using something in Graphs.jl?

[JeffBezanson]

The problem is that this means a given sparse matrix can change how its nzval vector is represented at different times, which is not desirable.

I think this patch is what we want:

diff --git a/base/sparse/sparsematrix.jl b/base/sparse/sparsematrix.jl
index b62ede3..0df0db7 100644
--- a/base/sparse/sparsematrix.jl
+++ b/base/sparse/sparsematrix.jl
@@ -389,7 +389,7 @@ function sprand(m::Integer, n::Integer, density::FloatingPoint, rng::Function, v
     if !iseltype(v,Bool)
         return sparse_IJ_sorted!(I, J, rng(length(uind)), m, n, +)  # it will never need to combine
     else
-        return sparse_IJ_sorted!(I, J, trues(length(uind)), m, n, +)
+        return sparse_IJ_sorted!(I, J, ones(Bool,length(uind)), m, n, +)
     end
 end

[timholy]

@jperla, the JIT can help only if the types can be determined at runtime. The issues here are discussed at length in the FAQ: http://docs.julialang.org/en/latest/manual/faq/#how-do-abstract-or-ambiguous-fields-in-types-interact-with-the-compiler and the next section, which deals directly with the issue here.

For a better design, see JuliaLang/LinearAlgebra.jl#84

[jperla]

@JeffBezanson okay i updated the code to use ones(), all tests pass. Thanks!

[jperla]

thanks @timholy

[ViralBShah]

I think that the sprand API should be revisited and made similar to the other recent changes to rand. Also, we probably can do without sprandbool.