reducedims for arrays with abstract element types. fix JuliaLang#28227

base/reducedim.jl

@@ -141,7 +141,9 @@ for (f1, f2, initval) in ((:min, :max, :Inf), (:max, :min, :(-Inf)))
  # but NaNs need to be avoided as intial values
  v0 = v0 != v0 ? typeof(v0)($initval) : v0
 
- return reducedim_initarray(A, region, v0)
+ T = promote_union(eltype(A))
+ Tr = v0 isa T ? T : typeof(v0)
+ return reducedim_initarray(A, region, v0, Tr)
  end
  end
 end

stdlib/SparseArrays/src/sparsematrix.jl

@@ -1613,9 +1613,10 @@ end
 
 # In general, output of sparse matrix reductions will not be sparse,
 # and computing reductions along columns into SparseMatrixCSC is
-# non-trivial, so use Arrays for output
-Base.reducedim_initarray(A::SparseMatrixCSC, region, v0, ::Type{R}) where {R} =
- fill(v0, Base.reduced_indices(A,region))
+# non-trivial, so use Arrays for output. Array element type is given by `R`.
+function Base.reducedim_initarray(A::SparseMatrixCSC, region, v0, ::Type{R}) where {R}
+ fill!(Array{R}(undef, Base.to_shape(Base.reduced_indices(A, region))), v0)
+end
 
 # General mapreduce
 function _mapreducezeros(f, op, ::Type{T}, nzeros::Int, v0) where T

stdlib/SparseArrays/test/sparse.jl

@@ -523,8 +523,9 @@ end
 
 @testset "reductions" begin
  pA = sparse(rand(3, 7))
+ p28227 = sparse(Real[0 0.5])
 
- for arr in (se33, sA, pA)
+ for arr in (se33, sA, pA, p28227)
  for f in (sum, prod, minimum, maximum)
  farr = Array(arr)
  @test f(arr) ≈ f(farr)