Fix sort with unused levels (#273)

The code did not handle correctly the case when unused levels were present at the front.
This seems to have become easier to trigger since the index vs. levels distinction
was dropped.

src/array.jl

@@ -943,16 +943,17 @@ function Base.sort!(v::CategoricalVector;
 
     # compute the order in which to read from counts
     ord = Base.Sort.ord(lt, by, rev, order)
-    index = eltype(v) >: Missing ? [missing; v.pool.valindex] : v.pool.valindex
     seen = counts .> 0
     anymissing = eltype(v) >: Missing && seen[1]
-    perm = sortperm(view(index, seen), order=ord)
-    nzcounts = counts[seen]
+    levs = eltype(v) >: Missing ? [missing; v.pool.valindex] : v.pool.valindex
+    sortedlevs = sort!(Vector(view(levs, seen)), order=ord)
+    levelsmap = something.(indexin(sortedlevs, levs))
     j = 0
     refs = v.refs
-    @inbounds for ref in perm
-        tmpj = j + nzcounts[ref]
-        refs[(j+1):tmpj] .= ref - anymissing
+    @inbounds for i in eachindex(levelsmap)
+        ref = levelsmap[i]
+        tmpj = j + counts[ref]
+        refs[(j+1):tmpj] .= ref - (eltype(v) >: Missing)
         j = tmpj
     end
 

test/13_arraycommon.jl

@@ -926,7 +926,11 @@ end
             cv = categorical(v)
             levels!(cv, ["b", "a", "c", "d"])
             @test sort(cv, rev=rev) ≅
-                ["b", "a", "c", "d", missing][sort([5, 1, 2, 3, 4][cv.refs .+ 1], rev=rev)]
+                ["b", "a", "c", "d", missing][sort([5; 1:4][cv.refs .+ 1], rev=rev)]
+
+            levels!(cv, ["x", "z", "b", "a", "y", "c", "d", "0"])
+            @test sort(cv, rev=rev) ≅
+                ["x", "z", "b", "a", "y", "c", "d", "0", missing][sort([9; 1:8][cv.refs .+ 1], rev=rev)]
 
             cv = categorical(v)
             @test sort(cv, rev=rev, lt=(x, y) -> isless(y, x)) ≅