Fix a type-instability in sparse findmin/findmax

The helper function `_findr` would usually return a `Vector` as first
argument, but would use a `SparseMatrixCSC` in the empty case. Fix by
always using `Vector`.

Author: martinholters

stdlib/SparseArrays/src/sparsematrix.jl

@@ -2045,7 +2045,7 @@ function _findr(op, A, region, Tv)
             throw(ArgumentError("array slices must be non-empty"))
         else
             ri = Base.reduced_indices0(A, region)
-            return (similar(A, ri), zeros(Ti, ri))
+            return (zeros(Tv, ri), zeros(Ti, ri))
         end
     end
 

stdlib/SparseArrays/test/sparse.jl

@@ -1338,10 +1338,10 @@ end
 @testset "argmax, argmin, findmax, findmin" begin
     S = sprand(100,80, 0.5)
     A = Array(S)
-    @test argmax(S) == argmax(A)
-    @test argmin(S) == argmin(A)
-    @test findmin(S) == findmin(A)
-    @test findmax(S) == findmax(A)
+    @test @inferred(argmax(S)) == argmax(A)
+    @test @inferred(argmin(S)) == argmin(A)
+    @test @inferred(findmin(S)) == findmin(A)
+    @test @inferred(findmax(S)) == findmax(A)
     for region in [(1,), (2,), (1,2)], m in [findmax, findmin]
         @test m(S, dims=region) == m(A, dims=region)
     end